{ "schemaVersion": 4, "updatedAt": "2026-10-30T11:53:42+01:00", "mentalModel": { "layers": [ { "name": "Data & Knowledge", "description": "The information the AI can use: documents, databases, retrieval, context, memory and the systems that keep that information current." }, { "name": "Model", "description": "The model that performs language, reasoning, coding, vision or audio tasks. This is the raw capability layer." }, { "name": "Agent & Infrastructure", "description": "The runtime around the model: tools, state, memory, orchestration, permissions, sandboxes and the ability to take actions." }, { "name": "Application & Workflow", "description": "The user experience and business process where AI is used, from copilots to end-to-end task flows." }, { "name": "Production & Business", "description": "Everything needed to run AI responsibly at scale: evals, monitoring, governance, cost, risk, reliability and measurable business value." } ], "primitives": [ { "name": "Compute", "description": "Where it runs. This covers model inference, agent runtimes, application code, indexing, evals and other processing." }, { "name": "Networking", "description": "How it connects. This covers traffic between users, models, tools, data sources, APIs and enterprise systems." }, { "name": "Storage", "description": "Where data and state live. This includes documents, databases, vectors, memory, logs, traces and application state." }, { "name": "Security", "description": "Who or what is allowed access. This includes identity, permissions, secrets, isolation, approvals and audit controls." } ], "impacts": [ { "name": "Disruptive", "description": "Changes what is practical to build or reshapes the market enough that old assumptions need to be reconsidered." }, { "name": "Step-change", "description": "A clear jump in capability, cost, reliability or usability that changes architecture decisions, but does not reset the whole market." }, { "name": "Incremental", "description": "A meaningful improvement to an existing approach. It matters, but the basic architecture and workflow stay mostly the same." }, { "name": "Packaging / distribution", "description": "The underlying capability mostly existed already, but it becomes easier, cheaper, more integrated or available to more teams." } ] }, "trajectories": [ { "layer": "Data & Knowledge", "description": "Getting the right information into the system.", "now": "Long context, retrieval, connected company data, citations, and memory are increasingly combined.", "direction": "Permission-aware organizational context that stays fresh and can be reused across models and agents.", "constraint": "Freshness, provenance, permissions, data quality, and cost." }, { "layer": "Model", "description": "Raw reasoning and multimodal capability.", "now": "Frontier models reason longer, operate computers, work across modalities, and improve rapidly on professional tasks.", "direction": "Broader, more general capability with adaptive reasoning and model routing by task, cost, latency, and risk.", "constraint": "Alignment, inference cost, reliability, misuse risk, and knowing when capability is actually trustworthy." }, { "layer": "Agent & Infrastructure", "description": "Turning model capability into sustained action.", "now": "Managed runtimes, tools, MCP, subagents, computer use, state, and tracing support longer tasks.", "direction": "Persistent software actors that can work for hours or days, recover from failure, coordinate, and operate across systems.", "constraint": "Long-horizon reliability, identity, bounded permissions, verification, and recovery." }, { "layer": "Application & Workflow", "description": "Where AI changes how work is actually done.", "now": "The pattern is moving from assistance beside the workflow to agents executing meaningful parts of the workflow.", "direction": "AI-native processes where machines execute routine and analytical work while people handle goals, judgment, approvals, and exceptions.", "constraint": "Workflow redesign, adoption, accountability, process ownership, and proving business value." }, { "layer": "Production & Business", "description": "Making AI controllable, measurable, and economically useful. LLMOps and GenAIOps practices for running GenAI in production grew gradually from 2023 onward.", "now": "Evals, traces, trajectory monitoring, governance, approval gates, and cost controls are becoming core production architecture.", "direction": "Continuous risk-based control where capability, alignment, reliability, cost, and business value are measured together.", "constraint": "Scalable alignment and oversight as models become more capable and autonomous." } ], "updates": [ { "id": "2022-11-30-chatgpt", "date": "2022-11-30", "layer": "Application & Workflow", "kind": "milestone", "changeType": "interface shift", "impact": "Disruptive", "title": "OpenAI releases ChatGPT as a public research preview", "whatChanged": "OpenAI released ChatGPT as a free research preview. People could talk to a language model in plain conversation and ask follow-up questions.", "whyCare": "Chat became the way most people first used a language model.", "before": "People used language models mainly through research demos, APIs and specialist tools.", "after": "Anyone with a browser could use a capable model through a chat window.", "learn": "Treat chat as one interface among several. Keep the conversation layer separate from the model and the business logic behind it.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - Introducing ChatGPT", "url": "https://openai.com/index/chatgpt/" } ], "architectureImplication": "Design the chat interface, the model and the business workflow as separate layers, so each can change on its own.", "tradeoffs": [ "Reliability", "Control", "Cost" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [], "short": "ChatGPT" }, { "id": "2023-03-14-gpt4", "date": "2023-03-14", "layer": "Model", "kind": "milestone", "changeType": "model capability", "impact": "Step-change", "title": "OpenAI releases GPT-4, which accepts both images and text", "whatChanged": "OpenAI released GPT-4, a large multimodal model that accepts image and text input. It improved on professional and academic benchmarks.", "whyCare": "Harder knowledge work and image-plus-text use cases became realistic, and so did the need for evaluation and controls.", "before": "Most GenAI applications ran on weaker, text-only models.", "after": "Teams could build for more complex reasoning, higher-stakes workflows and image-plus-text input.", "learn": "Test the model on your own tasks before you raise the stakes. Add evaluation and controls as capability grows.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - GPT-4 research", "url": "https://openai.com/index/gpt-4-research/" }, { "name": "OpenAI - GPT-4", "url": "https://openai.com/index/gpt-4/" } ], "architectureImplication": "A stronger model still needs evaluations, good data, controls and workflow design around it.", "tradeoffs": [ "Accuracy", "Latency", "Cost" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "model-selection" ], "short": "GPT-4" }, { "id": "2023-06-13-function-calling", "date": "2023-06-13", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "tool use", "impact": "Step-change", "title": "OpenAI adds function calling so models can return structured tool calls", "whatChanged": "OpenAI added function calling to its API. A model can pick a described function and return structured arguments that code can use to call an API or tool.", "whyCare": "Applications could let models trigger software actions reliably.", "before": "Developers parsed free text or relied on fragile prompt conventions to connect models to software.", "after": "Models return structured tool calls that code can validate and run.", "learn": "Let the model choose the action and let your code run it. Check arguments and permissions before any tool runs.", "cloudPrimitives": [ "Compute", "Networking", "Security" ], "sources": [ { "name": "OpenAI - Function calling and other API updates", "url": "https://openai.com/index/function-calling-and-other-api-updates/" } ], "architectureImplication": "Separate model judgment from deterministic execution. Put validation and permission checks between them.", "tradeoffs": [ "Reliability", "Security", "Control" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "agent-vs-workflow", "integration-standards" ], "short": "Function calling" }, { "id": "2024-02-15-gemini15-context", "date": "2024-02-15", "layer": "Data & Knowledge", "kind": "milestone", "changeType": "context expansion", "impact": "Step-change", "title": "Google previews Gemini 1.5 Pro with a one-million-token context window", "whatChanged": "Google introduced Gemini 1.5 Pro with a standard 128,000-token context window. A one-million-token window was available in private preview for selected developers and enterprise customers.", "whyCare": "Very long context changed the choice between retrieval and putting source material straight into the prompt.", "before": "Teams split large document sets into small chunks and retrieved a few at a time.", "after": "Some workloads could send far larger collections directly. Retrieval still mattered for cost, freshness, permissions and precision.", "learn": "Test long context, retrieval and a mix of both on your data. Compare cost, latency and accuracy.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Google - Introducing Gemini 1.5", "url": "https://blog.google/innovation-and-ai/products/google-gemini-next-generation-model-february-2024/" } ], "architectureImplication": "Choose between long context, retrieval or both based on data size, freshness, latency, cost and access control.", "tradeoffs": [ "Accuracy", "Security", "Cost" ], "verified": "2026-10-17", "vendors": [ "Google" ], "decisions": [ "rag-vs-long-context" ], "short": "1M context" }, { "id": "2024-05-13-gpt4o", "date": "2024-05-13", "layer": "Model", "kind": "milestone", "changeType": "multimodal interaction", "impact": "Step-change", "title": "OpenAI releases GPT-4o, one model for text, vision and audio", "whatChanged": "OpenAI introduced GPT-4o, trained as a single model across text, vision and audio. It can respond to audio input in as little as 232 milliseconds.", "whyCare": "Voice and multimodal products need fewer separate speech, vision and language components.", "before": "Voice applications chained several models and conversion steps.", "after": "One model can handle more of the interaction directly.", "learn": "For voice and multimodal products, plan for streaming, interruptions, session state and cost.", "cloudPrimitives": [ "Compute", "Networking", "Storage" ], "sources": [ { "name": "OpenAI - Hello GPT-4o", "url": "https://openai.com/index/hello-gpt-4o/" } ], "architectureImplication": "Design multimodal systems around streaming, latency, session state and interruption handling.", "tradeoffs": [ "Accuracy", "Latency", "Cost" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "model-selection" ], "short": "GPT-4o" }, { "id": "2024-07-23-llama31", "date": "2024-07-23", "layer": "Model", "kind": "milestone", "changeType": "open model ecosystem", "impact": "Step-change", "title": "Meta releases Llama 3.1, including an openly available 405B model", "whatChanged": "Meta released Llama 3.1, including a 405B model with a 128K context window. The weights are available for broad use under Meta's license.", "whyCare": "Teams could run a model close to the frontier on infrastructure they control.", "before": "Frontier-level capability sat mostly behind closed, hosted APIs.", "after": "Enterprises had a strong open-weight option for control, customization and deployment.", "learn": "Compare licensing, data location, hosting cost and the skills you need to run it before you self-host.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Meta - Introducing Llama 3.1", "url": "https://ai.meta.com/blog/meta-llama-3-1/" } ], "architectureImplication": "Weigh who runs the model, the license and the effort to operate it, next to quality and price.", "tradeoffs": [ "Accuracy", "Latency", "Cost" ], "verified": "2026-10-17", "vendors": [ "Meta", "Open models" ], "decisions": [ "hosted-vs-open" ], "short": "Llama 3.1" }, { "id": "2024-09-12-o1", "date": "2024-09-12", "layer": "Model", "kind": "milestone", "changeType": "reasoning model", "impact": "Step-change", "title": "OpenAI releases o1, a model that reasons before it answers", "whatChanged": "OpenAI released o1-preview, a model series that spends more time reasoning before responding. It improves on harder science, coding and math tasks.", "whyCare": "Reasoning time became a setting you choose, with a trade-off between quality, latency and cost.", "before": "Teams treated most model calls as the same kind of single-pass answer.", "after": "Teams can choose when a task is worth more reasoning time and cost.", "learn": "Route tasks by difficulty and value. Save expensive reasoning for the tasks that need it.", "cloudPrimitives": [ "Compute", "Networking" ], "sources": [ { "name": "OpenAI - Introducing o1-preview", "url": "https://openai.com/index/introducing-openai-o1-preview/" } ], "architectureImplication": "Add routing between fast and reasoning models, based on task difficulty, risk and cost.", "tradeoffs": [ "Accuracy", "Latency", "Cost" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "model-selection", "cost-and-routing" ], "short": "o1" }, { "id": "2024-10-22-computer-use", "date": "2024-10-22", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "computer interaction", "impact": "Step-change", "title": "Anthropic releases computer use, letting Claude operate graphical software", "whatChanged": "Anthropic released computer use in public beta. Claude can look at screenshots, move a pointer, click and type through software interfaces.", "whyCare": "Agents can act in systems that have no API, though reliability and security remain hard.", "before": "Automation needed APIs, scripts or fragile browser automation.", "after": "A model can operate existing graphical software in a general way.", "learn": "Run computer-use agents in a sandbox. Limit credentials and plan how to check and undo their actions.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Anthropic - Developing a computer use model", "url": "https://www.anthropic.com/news/developing-computer-use" } ], "architectureImplication": "Computer use widens what agents can reach. Add sandboxing, credential limits, verification and recovery.", "tradeoffs": [ "Reliability", "Security", "Control" ], "verified": "2026-10-17", "vendors": [ "Anthropic" ], "decisions": [ "agent-vs-workflow", "security-and-guardrails" ], "short": "Computer use" }, { "id": "2024-11-25-mcp", "date": "2024-11-25", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "open protocol", "impact": "Step-change", "title": "Anthropic open-sources MCP, a standard for connecting AI to tools and data", "whatChanged": "Anthropic open-sourced the Model Context Protocol, a standard way to connect AI assistants to data sources, business tools and development environments.", "whyCare": "Teams can build a connector once and use it with many AI clients.", "before": "Each application or framework needed its own connection to each tool and data source.", "after": "Tools and data sources can expose one interface that many AI clients use.", "learn": "Check which of your tools already offer MCP servers. Adopt a protocol once the ecosystem around it grows.", "cloudPrimitives": [ "Networking", "Storage", "Security" ], "sources": [ { "name": "Anthropic - Introducing the Model Context Protocol", "url": "https://www.anthropic.com/news/model-context-protocol" } ], "architectureImplication": "Use a shared protocol for tool and data connections to reduce coupling between AI clients and systems.", "tradeoffs": [ "Reliability", "Security", "Control" ], "verified": "2026-10-17", "vendors": [ "Anthropic", "Open standards" ], "decisions": [ "integration-standards" ], "short": "MCP" }, { "id": "2025-01-20-deepseek-r1", "date": "2025-01-20", "layer": "Model", "kind": "milestone", "changeType": "open reasoning economics", "impact": "Step-change", "title": "DeepSeek releases R1, an open-weight reasoning model with low API prices", "whatChanged": "DeepSeek released R1 and smaller distilled reasoning models with open weights under the MIT license. API prices were $0.14 to $0.55 per million input tokens and $2.19 per million output tokens.", "whyCare": "Reasoning became available at lower cost and with open weights, which put pressure on closed-model prices.", "before": "Advanced reasoning came mainly from a few closed providers.", "after": "Teams had more choice across quality, hosting control, openness and cost.", "learn": "Keep your design portable across providers. Test new models quickly when capability and prices move.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "DeepSeek - R1 release", "url": "https://deepseek.com/en/news/deepseek-r1/" } ], "architectureImplication": "Avoid tying the architecture to one provider. Model routing and portability gain value as prices change.", "tradeoffs": [ "Accuracy", "Latency", "Cost" ], "verified": "2026-10-17", "vendors": [ "DeepSeek", "Open models" ], "decisions": [ "hosted-vs-open", "cost-and-routing" ], "short": "DeepSeek-R1" }, { "id": "2025-03-11-responses-agents", "date": "2025-03-11", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "managed agent platform", "impact": "Packaging / distribution", "title": "OpenAI launches the Responses API, built-in tools and an Agents SDK", "whatChanged": "OpenAI released the Responses API, built-in tools for web search, file search and computer use, and an Agents SDK with tracing.", "whyCare": "Teams could buy core agent infrastructure as a managed service and write less orchestration code.", "before": "Teams assembled tool use, orchestration, tracing and state with custom code or separate frameworks.", "after": "More of the agent harness comes from the platform.", "learn": "Decide which parts of the agent harness to buy and which to build. Check what you can export or move later.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - New tools for building agents", "url": "https://openai.com/index/new-tools-for-building-agents/" } ], "architectureImplication": "Managed agent runtimes change build-versus-buy decisions. Weigh speed of delivery against portability.", "tradeoffs": [ "Reliability", "Security", "Control" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "managed-vs-custom-runtime", "build-vs-buy" ], "short": "Agents SDK" }, { "id": "2025-04-09-a2a", "date": "2025-04-09", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "agent interoperability", "impact": "Incremental", "title": "Google announces A2A, a protocol for agents to work with other agents", "whatChanged": "Google announced the Agent2Agent protocol so agents from different vendors and frameworks can communicate and coordinate.", "whyCare": "Agent-to-agent interoperability became a stated industry goal. Standards still need adoption before teams can rely on them.", "before": "Multi-agent systems assumed one framework, one vendor or custom conventions.", "after": "A shared protocol could reduce coupling between agent systems as adoption grows.", "learn": "Track which vendors support A2A. Adopt it when your partners or customers need cross-vendor agents.", "cloudPrimitives": [ "Networking", "Security" ], "sources": [ { "name": "Google - Announcing the Agent2Agent Protocol", "url": "https://developers.googleblog.com/a2a-a-new-era-of-agent-interoperability/" } ], "architectureImplication": "Plan agent boundaries so a standard protocol can replace custom connections later.", "tradeoffs": [ "Reliability", "Security", "Control" ], "verified": "2026-10-17", "vendors": [ "Google", "Open standards" ], "decisions": [ "integration-standards" ], "short": "A2A" }, { "id": "2025-05-21-managed-mcp-background", "date": "2025-05-21", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "managed runtime maturity", "impact": "Incremental", "title": "OpenAI adds remote MCP servers and background mode to the Responses API", "whatChanged": "OpenAI added support for remote MCP servers, a background mode for long-running tasks, and further reliability and privacy features to its Responses API.", "whyCare": "Managed runtimes now handle more of the connections and long-running execution that agents need.", "before": "Teams built connectivity and long-running execution for agents themselves.", "after": "The platform covers more of the operational agent harness.", "learn": "Check what the managed runtime lets you see, export and secure. Decide how much lock-in you accept for the convenience.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - New tools and features in the Responses API", "url": "https://openai.com/index/new-tools-and-features-in-the-responses-api/" } ], "architectureImplication": "Managed runtimes can absorb connectivity and long-running work. Keep observability and an exit path in your design.", "tradeoffs": [ "Reliability", "Security", "Control" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "managed-vs-custom-runtime", "integration-standards" ], "short": "Remote MCP" }, { "id": "2026-09-18-long-running-agent-reliability", "date": "2026-09-18", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "architecture frontier", "impact": "Step-change", "title": "Microsoft Research shows agents get less reliable on longer tasks", "whatChanged": "Microsoft Research shows that frontier agents can already do multi-step work, but they get less reliable as tasks get longer and tool use gets more complex. Engineering work is moving to state, recovery, checks and control over the whole run.", "whyCare": "The question for enterprise teams moves from whether an agent can do one task to whether it can finish a long workflow safely, again and again.", "before": "Teams judged agents mainly on single tasks and demos.", "after": "Teams design for reliable runs across many steps, tools, failures and approvals.", "learn": "Keep agent state outside the model. Add retries, recovery points and checks between steps, and hand the run to a person when it goes wrong.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Microsoft Research - long-horizon tasks", "url": "https://www.microsoft.com/en-us/research/publication/on-training-large-language-models-for-long-horizon-tasks-an-empirical-study-of-horizon-length/" } ], "architectureImplication": "Design long agent runs like workflows: save state, recover from failures, check each step and escalate to people.", "tradeoffs": [ "Reliability", "Security", "Control" ], "vendors": [ "Microsoft" ], "decisions": [ "observability-and-evals", "state-and-memory" ] }, { "id": "2026-09-18-agent-security-controls", "date": "2026-09-18", "layer": "Production & Business", "kind": "signal", "changeType": "operating model", "impact": "Step-change", "title": "Each agent needs its own identity, permissions and audit trail", "whatChanged": "As agents get more tools and more autonomy, enterprises need to give each one its own identity, scoped permissions, a sandbox, audit logs and approval gates.", "whyCare": "An agent with access to company systems acts like a software user. It needs its own operating controls.", "before": "Teams secured the AI application and its API access.", "after": "Teams secure each agent as an actor with its own identity, permissions and logged actions.", "learn": "Give each agent its own credentials with the fewest permissions it needs. Run it in a sandbox, log what it does and require approval for risky actions.", "cloudPrimitives": [ "Networking", "Storage", "Security" ], "sources": [ { "name": "Google Cloud - workload identities", "url": "https://cloud.google.com/iam/docs/workload-identities" } ], "architectureImplication": "Treat every agent as a separate actor in your identity and access design, with its own credentials, limits and audit trail.", "tradeoffs": [ "Reliability", "Security", "Control" ], "vendors": [ "Google Cloud" ], "decisions": [ "identity-and-permissions", "security-and-guardrails" ] }, { "id": "2026-09-10-managed-agent-runtimes", "date": "2026-09-10", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "packaging / distribution", "impact": "Packaging / distribution", "title": "More providers package agent runtimes as managed services", "whatChanged": "More providers package context, tools, subagents, execution and long-running task support into managed agent runtimes.", "whyCare": "Teams could already build these pieces. Managed versions cut the work, which changes the build or buy decision.", "before": "Teams built most of the agent harness themselves or with separate frameworks.", "after": "Teams can use more of the runtime as managed infrastructure.", "learn": "List what your agent harness does today: state, context, tools, execution, retries, permissions, memory and monitoring. Check which parts a managed runtime covers.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - agent platform reference", "url": "https://openai.com/index/new-tools-for-building-agents/" } ], "architectureImplication": "Decide part by part what to build and what to buy. Keep state and memory portable so you can switch runtimes later.", "tradeoffs": [ "Reliability", "Security", "Control" ], "vendors": [ "OpenAI" ], "decisions": [ "managed-vs-custom-runtime" ] }, { "id": "2026-09-01-enterprise-evals", "date": "2026-09-01", "layer": "Production & Business", "kind": "signal", "changeType": "evaluation practice", "impact": "Incremental", "title": "Teams now check the path an agent takes as well as its answer", "whatChanged": "Teams now need to check which tools an agent used, which path it took and whether it stayed inside policy. Checking the final answer alone misses these.", "whyCare": "A correct-looking answer can come from an unsafe path. This matters once agents take actions.", "before": "Teams evaluated whether the final answer was correct.", "after": "Teams evaluate the result and the sequence of actions that produced it.", "learn": "Add trajectory checks to your evaluations. Keep audit logs and use human review for risky actions.", "cloudPrimitives": [ "Compute", "Storage", "Security" ], "sources": [ { "name": "Microsoft Research - diagnosing agent failures", "url": "https://www.microsoft.com/en-us/research/publication/diagnosing-with-insights-structured-analysis-of-agent-failures-via-behavioral-abstractions/" } ], "architectureImplication": "Combine output evaluations, trajectory evaluations, audit logs, policy checks and risk-based human review.", "tradeoffs": [ "Reliability", "Security", "Control" ], "vendors": [ "Microsoft" ], "decisions": [ "observability-and-evals" ] }, { "id": "2026-04-24-gemini-agent-platform", "date": "2026-04-24", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "enterprise agent platform", "impact": "Packaging / distribution", "title": "Google Cloud brings agent building, governance and scale into one platform", "whatChanged": "At Google Cloud Next 2026, Google Cloud introduced an end-to-end enterprise agent platform that covers building agents, governing them and running them at scale.", "whyCare": "Cloud providers now sell agent infrastructure as one platform, so teams can buy it whole.", "before": "Teams connected models, tools, identity and controls from separate services themselves.", "after": "Cloud platforms package these parts into one agent platform.", "learn": "Compare agent platforms on how easily you can move off them, what you can see and control, what they cost and how well they fit your stack.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Google Cloud Next 2026 recap", "url": "https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/google-cloud-next-26-recap/" } ], "architectureImplication": "An integrated platform saves build time but ties more of your stack to one vendor. Keep agent definitions and data portable.", "tradeoffs": [ "Reliability", "Security", "Control" ], "vendors": [ "Google Cloud", "Google" ], "decisions": [ "managed-vs-custom-runtime", "build-vs-buy" ] }, { "id": "2026-05-19-gemini-spark", "date": "2026-05-19", "layer": "Application & Workflow", "kind": "milestone", "changeType": "embedded agent workflow", "impact": "Step-change", "title": "Google announces Gemini Spark, a persistent agent across Workspace apps", "whatChanged": "Google announced Gemini Spark, a persistent personal agent that acts across Workspace applications under user direction.", "whyCare": "AI at work is moving from a chat tool next to the work to an agent inside the software people already use.", "before": "AI helped in single prompts or inside one application at a time.", "after": "A persistent agent can coordinate work across several everyday applications.", "learn": "Define what users may delegate, which actions need approval and what context the agent may see.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Google Workspace updates 2026", "url": "https://blog.google/products-and-platforms/products/workspace/workspace-updates/" } ], "architectureImplication": "Design for delegation, approvals, context boundaries and user control when agents act across applications.", "tradeoffs": [ "Reliability", "Control", "Cost" ], "verified": "2026-10-17", "vendors": [ "Google" ], "decisions": [ "human-approval" ], "short": "Gemini Spark" }, { "id": "2026-06-09-fable5", "date": "2026-06-09", "layer": "Model", "kind": "milestone", "changeType": "long-horizon model capability", "impact": "Step-change", "title": "Anthropic releases Claude Fable 5, built for long and complex tasks", "whatChanged": "Anthropic released Claude Fable 5. Anthropic says its lead over its other models grows with the length and complexity of the task, across coding, knowledge work, vision and research.", "whyCare": "How long and complex a task a model can finish is becoming a main way to compare models.", "before": "Hard tasks had to be split into many small steps with a person checking each one.", "after": "Stronger models can carry larger projects further before handing control back.", "learn": "Test models on how long and how complex a task they finish reliably. Use your own tasks.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Anthropic - Claude Fable 5", "url": "https://www.anthropic.com/news/claude-fable-5-mythos-5" } ], "architectureImplication": "Evaluate models on task length and complexity as well as benchmark scores.", "tradeoffs": [ "Accuracy", "Latency", "Cost" ], "verified": "2026-10-17", "vendors": [ "Anthropic" ], "decisions": [ "model-selection" ], "short": "Claude Fable 5" }, { "id": "2026-07-20-long-horizon-safety", "date": "2026-07-20", "layer": "Production & Business", "kind": "milestone", "changeType": "long-horizon safety", "impact": "Step-change", "title": "OpenAI moves safety evaluation from single actions to whole trajectories", "whatChanged": "OpenAI reported new failure types during limited internal use of a long-running model. It responded with new evaluations, long-horizon alignment work, trajectory monitoring and more user control.", "whyCare": "Once systems act over long periods, checking single outputs misses what the sequence of actions is working toward.", "before": "Safety checks focused on single responses or short, bounded tasks.", "after": "Teams need to judge whole sequences of actions and step in during a run.", "learn": "Add trajectory monitoring, long-horizon evaluations, approval points and recovery steps to agent designs.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - Safety and alignment in an era of long-horizon models", "url": "https://openai.com/index/safety-alignment-long-horizon-models/" } ], "architectureImplication": "Monitor whole agent runs. Define where a run must stop for approval and how to recover it.", "tradeoffs": [ "Reliability", "Security", "Control" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "observability-and-evals", "human-approval" ], "short": "Trajectory safety" }, { "id": "2026-09-01-workflow-execution", "date": "2026-09-01", "layer": "Application & Workflow", "kind": "signal", "changeType": "operating model", "impact": "Step-change", "title": "OpenAI examples show agents doing multi-step work inside company workflows", "whatChanged": "Published company examples show agents connected to tools and company context for onboarding, account management, developer work and other multi-step processes.", "whyCare": "The change that matters is delegation. Companies hand agents real parts of a business process.", "before": "AI mostly helped people draft, search and do single tasks beside the workflow.", "after": "Leading teams connect agents to context and tools and make the workflows that succeed repeatable.", "learn": "Start each agent workflow by defining the task boundary, the data and tools it needs, who approves what, what happens on failure and how you measure success.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - AI-native company workflows", "url": "https://openai.com/index/ai-native-company-workflows/" } ], "architectureImplication": "Design agent workflows like business processes, with clear inputs, approvals, failure handling and a measured outcome.", "tradeoffs": [ "Reliability", "Control", "Cost" ], "vendors": [ "OpenAI" ], "decisions": [ "agent-vs-workflow" ] }, { "id": "2026-09-03-gpt6-astra", "date": "2026-09-03", "layer": "Model", "kind": "milestone", "changeType": "frontier capability", "impact": "Step-change", "title": "OpenAI releases GPT-6 Astra with gains in computer use, science and cyber", "whatChanged": "OpenAI released GPT-6 Astra and reported state-of-the-art results on benchmarks for computer use, science, cybersecurity and professional work.", "whyCare": "Frontier models now operate software and carry out complex professional tasks as well as answering questions.", "before": "Models were less reliable at long, tool-using professional work and computer use.", "after": "A frontier model handles a wider set of complex tasks and operates software better.", "learn": "Test it on your own tool-using tasks. Treat vendor benchmark results as claims until you reproduce them.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - GPT-6 Astra", "url": "https://openai.com/index/gpt-6-astra/" }, { "name": "OpenAI - GPT-6 Astra safety overview", "url": "https://openai.com/index/safety-overview-gpt-6-astra/" } ], "architectureImplication": "Broader capability needs stronger controls, routing, monitoring and risk assessment around the model.", "tradeoffs": [ "Accuracy", "Latency", "Cost" ], "verified": "2026-10-17", "verificationNote": "The OpenAI page shows only a 29 September 2026 update. Launch date confirmed by news coverage.", "vendors": [ "OpenAI" ], "decisions": [ "model-selection" ], "short": "GPT-6 Astra" }, { "id": "2026-09-06-automated-research", "date": "2026-09-06", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "automated research", "impact": "Step-change", "title": "OpenAI says it now has an automated research intern under human supervision", "whatChanged": "OpenAI said coding agents were speeding up its internal research and that it had reached its goal of an automated research intern that works under human supervision.", "whyCare": "If AI speeds up AI research, new model capabilities may arrive faster, and so may the need for alignment and safety work.", "before": "AI systems mainly helped researchers with bounded tasks.", "after": "Research agents can take on larger experiment and coding loops under supervision.", "learn": "Expect model releases to keep speeding up. Plan for model changes as a regular event in your architecture.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - Research acceleration", "url": "https://openai.com/index/research-acceleration-view-inside-openai/" } ], "architectureImplication": "Build so you can swap models often. Keep evaluations ready and avoid tying business logic to one model's quirks.", "tradeoffs": [ "Reliability", "Security", "Control" ], "vendors": [ "OpenAI" ], "decisions": [] }, { "id": "2026-09-10-data-agent", "date": "2026-09-10", "layer": "Data & Knowledge", "kind": "signal", "changeType": "connected business data", "impact": "Packaging / distribution", "title": "OpenAI's data agent combines company data, analysis and dashboards in one workflow", "whatChanged": "OpenAI described a business data agent that connects to company data, investigates questions, runs analysis and builds interactive dashboards in one workflow.", "whyCare": "People without analyst skills can get from a question to a shareable analysis themselves.", "before": "Business users waited for analysts or moved by hand between BI tools, SQL, spreadsheets and reports.", "after": "A conversational agent can investigate connected business data and create shareable analysis directly.", "learn": "Before you connect such an agent, check data permissions, shared metric definitions, where each number comes from, how fresh it is and how you audit it.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - Now everyone can put data to work", "url": "https://openai.com/index/put-data-to-work/" } ], "architectureImplication": "The hard design work moves to the data layer: permissions, agreed metric definitions, lineage, freshness and audit logs.", "tradeoffs": [ "Accuracy", "Security", "Cost" ], "vendors": [ "OpenAI" ], "decisions": [ "build-vs-buy" ] }, { "id": "2026-09-10-financial-services", "date": "2026-09-10", "layer": "Application & Workflow", "kind": "signal", "changeType": "vertical workflow product", "impact": "Packaging / distribution", "title": "OpenAI launches ChatGPT for financial services with premium market data", "whatChanged": "OpenAI launched a financial-services edition of ChatGPT. It combines premium market data from partners such as Daloopa and PitchBook with GPT-6 Astra for research, modeling and client materials.", "whyCare": "AI products are becoming industry-specific, with data rights and workflows built in.", "before": "Teams combined a general assistant, data subscriptions, templates and controls themselves.", "after": "A vertical product packages trusted data, domain workflows, output formats and governance together.", "learn": "When you compare vertical AI products, check data rights, workflow fit and governance as closely as the model.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - ChatGPT for Financial Services", "url": "https://openai.com/index/introducing-chatgpt-financial-services/" } ], "architectureImplication": "Expect more industry products where data rights and workflow design matter as much as the model.", "tradeoffs": [ "Reliability", "Control", "Cost" ], "vendors": [ "OpenAI" ], "decisions": [ "build-vs-buy" ] }, { "id": "2026-09-16-work-expansion", "date": "2026-09-16", "layer": "Application & Workflow", "kind": "signal", "changeType": "work redesign", "impact": "Incremental", "title": "Research finds AI use adds new recurring tasks to people's jobs", "whatChanged": "Economic research reported that some AI-assisted tasks outside a worker's usual occupation become recurring parts of their work.", "whyCare": "AI may change what jobs include as well as how fast people work.", "before": "Teams measured AI adoption as faster work on the same tasks.", "after": "Some workers add new recurring tasks because AI lowers the skill or effort needed.", "learn": "Measure changes in tasks and roles alongside time saved.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "OpenAI - How workers are unlocking new ways of working", "url": "https://openai.com/index/unlocking-new-ways-of-working/" } ], "architectureImplication": "Measure workflow change, task expansion and role redesign alongside time savings.", "tradeoffs": [ "Reliability", "Control", "Cost" ], "vendors": [ "OpenAI" ], "decisions": [] }, { "id": "2026-09-18-agentcore-runtime-v2", "date": "2026-09-18", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "agent runtime", "impact": "Incremental", "title": "AWS AgentCore Runtime V2 starts agents faster and frees idle memory", "whatChanged": "AWS introduced AgentCore Runtime V2 with memory reclamation during sessions and snapshot-based startup. AWS measured about 2-second P75 cold starts across container images from 200 MB to 2 GB, versus roughly 5.4 to 30 seconds for the earlier runtime.", "whyCare": "Production agents increasingly need isolated compute that can sit idle, resume quickly, run for long periods, and avoid paying for peak memory throughout a session.", "before": "Teams often kept agent environments warm, paid for memory held at the session high-water mark, or built their own lifecycle machinery to reduce cold starts and idle cost.", "after": "A managed agent runtime can scale to zero, reclaim unused memory during a session, and restore prepared environments from snapshots with more predictable startup.", "learn": "Separate the agent harness from the agent runtime. The harness manages reasoning, tools and context; the runtime provides compute, isolation, state and lifecycle.", "architectureImplication": "Treat agent execution as its own layer. Compare managed runtimes on isolation, startup time, state, identity handling, how easily you can leave, and cost.", "tradeoffs": [ "Latency", "Cost", "Control" ], "cloudPrimitives": [ "Compute", "Storage", "Security" ], "sources": [ { "name": "AWS - The new AgentCore runtime", "url": "https://aws.amazon.com/blogs/machine-learning/the-new-agentcore-runtime-elastic-optimized-and-consistently-fast-starts/" }, { "name": "AWS - AgentCore Runtime V2 documentation", "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-how-it-works.html" } ], "vendors": [ "AWS" ], "decisions": [ "managed-vs-custom-runtime" ] }, { "id": "2026-09-18-hyperpod-inference-gateway", "date": "2026-09-18", "layer": "Production & Business", "kind": "signal", "changeType": "inference routing", "impact": "Packaging / distribution", "title": "AWS HyperPod Inference Gateway routes LLM requests by cache and queue state", "whatChanged": "AWS introduced SageMaker HyperPod Inference Gateway, which routes LLM requests using signals such as KV-cache utilization, queue depth, LoRA adapter residency, prefix-cache hit rate and running requests. AWS reports up to 82% lower time-to-first-token in its tests.", "whyCare": "LLM replicas are not interchangeable in the way ordinary stateless web servers are. Cache state, queues and loaded adapters can materially change latency and GPU efficiency.", "before": "Generic round-robin or least-connections routing could send a request to a model pod with a poor cache state, a deep queue or the wrong adapter loaded.", "after": "An inference-aware router can select a backend using model-serving state, improving cache reuse, latency and GPU utilization without changing application clients.", "learn": "Understand KV cache, prefix caching, time-to-first-token, prefill versus decode, LoRA adapters and inference-aware routing. These increasingly determine self-hosted model economics.", "architectureImplication": "For dedicated GPU inference, plan routing and cache layout as part of your model-serving design. A generic load balancer ignores the cache and queue state that drive latency.", "tradeoffs": [ "Latency", "Cost", "Reliability" ], "cloudPrimitives": [ "Compute", "Networking", "Storage" ], "sources": [ { "name": "AWS - Introducing SageMaker HyperPod Inference Gateway", "url": "https://aws.amazon.com/blogs/machine-learning/introducing-amazon-sagemaker-hyperpod-inference-gateway/" }, { "name": "AWS - HyperPod Inference release notes", "url": "https://docs.aws.amazon.com/sagemaker/latest/dg/sagemaker-hyperpod-inference-release-notes.html" } ], "vendors": [ "AWS" ], "decisions": [ "cost-and-routing", "deployment" ] }, { "id": "2026-09-22-gpt-6-sol-luna-cost-capability", "date": "2026-09-22", "layer": "Model", "kind": "signal", "changeType": "model capability and economics", "impact": "Step-change", "title": "OpenAI releases GPT-6 Sol and Luna as lower-cost model tiers", "whatChanged": "OpenAI released GPT-6 Sol and Luna on 22 September. Per million tokens, Sol costs $2 input and $10 output, and Luna $0.10 input and $0.50 output. OpenAI says they bring GPT-6 Astra advances to cheaper tiers.", "whyCare": "More reasoning, coding and agent subtasks may become affordable at high volume. Compare task success, latency and total cost with retries included.", "before": "Teams kept strong models for hard steps and used older, cheaper models for high-volume steps.", "after": "Teams can test newer, cheaper tiers for both hard and high-volume steps.", "learn": "Build task-specific evaluations. Route by accuracy, latency, cache hit rate, retries and cost per successful task.", "architectureImplication": "Benchmark several models and route by workload and risk. Keep a provider abstraction and a fallback.", "tradeoffs": [ "Accuracy", "Latency", "Cost", "Reliability", "Control" ], "cloudPrimitives": [ "Compute", "Networking", "Security" ], "sources": [ { "name": "OpenAI - Introducing GPT-6 Sol and Luna", "url": "https://openai.com/index/introducing-gpt-6-sol-and-luna/" }, { "name": "OpenAI API changelog", "url": "https://developers.openai.com/api/docs/changelog" }, { "name": "OpenAI API pricing", "url": "https://developers.openai.com/api/docs/pricing" } ], "vendors": [ "OpenAI" ], "decisions": [ "cost-and-routing", "model-selection" ] }, { "id": "2026-09-22-claude-opus-55-agent-economics", "date": "2026-09-22", "layer": "Model", "kind": "signal", "changeType": "model pricing and migration", "impact": "Step-change", "title": "Anthropic releases Claude Opus 5.5 at lower prices than Opus 5", "whatChanged": "Anthropic released Claude Opus 5.5 on 22 September at $4 input and $20 output per million tokens, with cache reads at $0.20. Anthropic reports typical workloads cost 40% less than on Opus 5. It has a 1M-token context window and always-on adaptive thinking.", "whyCare": "Agents that reread cached context often may cost less to run. Upgrades need testing because several thinking and tool-use behaviours changed.", "before": "Opus 5 had higher token and cache-read prices, and existing integrations could assume optional thinking or forced tool use.", "after": "Opus 5.5 offers cheaper cached context and faster output, according to Anthropic. Integrations need checks for always-on thinking and tool-use changes.", "learn": "Measure total agent cost and cache hit rate on your own tasks. Run migration tests for thinking and tool-use behaviour.", "architectureImplication": "Test Opus 5.5 on representative agent tasks and run migration tests for thinking/tool-use behavior before changing production model routing.", "tradeoffs": [ "Accuracy", "Latency", "Cost", "Reliability", "Security", "Control" ], "cloudPrimitives": [ "Compute", "Storage", "Security" ], "sources": [ { "name": "Anthropic - Claude Opus 5.5 announcement", "url": "https://www.anthropic.com/claude-opus-5-5" }, { "name": "Anthropic - Opus 5.5 platform docs", "url": "https://platform.claude.com/docs/en/models/opus-5-5/overview" } ], "vendors": [ "Anthropic" ], "decisions": [ "cost-and-routing", "model-selection" ] }, { "id": "2026-09-23-microsoft-ai-work-system-economics", "date": "2026-09-23", "layer": "Production & Business", "kind": "signal", "changeType": "enterprise AI operating model", "impact": "Incremental", "title": "Microsoft describes an AI operating model with separate economics for agent work", "whatChanged": "Microsoft described an enterprise AI operating model with five parts. They are a choice of models, company data and context, central monitoring and governance, loops that keep improving agents, and separate economics for human subscriptions and agent work. Each part existed before. They now come together as a mainstream operating model.", "whyCare": "Agents do variable amounts of work, so a flat price per user no longer covers the cost. Teams need to measure cost and quality per finished task or workflow.", "before": "Companies bought enterprise AI mainly as a per-user copilot subscription. Agent cost, testing and controls were handled separately or as experiments.", "after": "Companies treat agents as metered workloads: they choose models, ground agents in company data, monitor them centrally, test them continuously and pay by use.", "learn": "Track cost per successful task. Learn how model routing and agent test loops work, and how seat pricing differs from paying for agent work by use.", "architectureImplication": "Build agent platforms that let you switch models and record usage, outcomes and cost per workflow or agent. Do not assume one model and one flat license.", "tradeoffs": [ "Cost", "Reliability", "Security", "Control" ], "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Microsoft - Building the system for AI at work", "url": "https://www.microsoft.com/en-us/copilot/blog/2026/09/23/building-the-system-for-ai-at-work/" } ], "vendors": [ "Microsoft" ], "decisions": [ "cost-and-routing" ] }, { "id": "2026-09-24-cross-platform-agent-management", "date": "2026-09-24", "layer": "Production & Business", "kind": "signal", "changeType": "agent governance", "impact": "Packaging / distribution", "title": "Dataiku launches Agent Management to govern agents across vendor platforms", "whatChanged": "Dataiku launched Agent Management as a standalone product. It finds and lists agents across major vendor platforms and custom OpenTelemetry setups. It monitors their health, usage, cost, quality and behavior, and keeps risk and certification records. The new part is a vendor-neutral layer above separate agent stacks.", "whyCare": "Companies that run agents on several clouds and SaaS platforms need one view of all of them. No single agent platform gives that view.", "before": "Teams monitored and governed each agent inside the platform where it was built, so inventory, ownership, cost, quality and risk were split across tools.", "after": "A separate layer can find agents across platforms, map their models and tools, track operating and business measures, and keep risk and certification evidence.", "learn": "Learn how agent inventories and OpenTelemetry-based monitoring work, and why running agents from several vendors needs one place to see and control them.", "architectureImplication": "If your agents run on several platforms, consider one governance layer above them. Each vendor's own monitoring only sees its own agents.", "tradeoffs": [ "Cost", "Reliability", "Security", "Control" ], "cloudPrimitives": [ "Networking", "Storage", "Security" ], "sources": [ { "name": "Dataiku - Agent Management general availability announcement", "url": "https://www.dataiku.com/company/news/dataiku-agent-management-general-availability" }, { "name": "Dataiku - Discover Agent Management", "url": "https://www.dataiku.com/blog/dataiku-agent-management" } ], "vendors": [ "Other" ], "decisions": [ "observability-and-evals", "identity-and-permissions" ] }, { "id": "2026-10-09-postman-agent-mode-production-patterns", "date": "2026-10-09", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "production agent architecture", "impact": "Incremental", "title": "Postman shows how it limits tools and context for its production agent", "whatChanged": "Postman and AWS described Agent Mode's design. A root agent narrows over 170 tools to about 15 per task. Schema-aware queries replace many read tools, and context handlers supply relevant state. State-changing actions still need approval. Tool-selection errors rose above about 40 visible tools in Postman's testing.", "whyCare": "Past a point, more tools and prompt text make agents less reliable. Shaping context and limiting actions keeps them working.", "before": "Early versions used many small tools and UI-shaped data, which raised errors and latency.", "after": "Tools are chosen per task, reads use schema-aware queries, and approvals gate changes.", "learn": "Count your tool catalog as part of the context budget. Select tools per task and keep approval for important writes.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "AWS - How Postman runs Agent Mode for 40 million developers on Amazon Bedrock", "url": "https://aws.amazon.com/blogs/machine-learning/how-postman-runs-agent-mode-for-40-million-developers-on-amazon-bedrock/" } ], "architectureImplication": "Give each task a small, relevant tool set and curated context. Keep write approvals, data retention and inference region explicit.", "tradeoffs": [ "Reliability", "Latency", "Cost", "Security", "Control" ], "vendors": [ "AWS", "Other" ], "decisions": [ "agent-vs-workflow", "human-approval" ] }, { "id": "2026-10-09-ai2-gpu-time-budget-scheduler", "date": "2026-10-09", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "GPU cluster scheduling", "impact": "Incremental", "title": "Ai2 replaces GPU priorities with time budgets and preemption contracts", "whatChanged": "Ai2 replaced its priority-based GPU scheduler with fair-share allocation weighted by GPU-time budgets, plus minimum-runtime contracts that let resumable jobs be preempted. In 30 days it delivered 98% of allocated GPU hours at 98% occupancy. Debug-job p90 waits fell from two hours to 30 seconds.", "whyCare": "Scarce GPUs tempt teams to hold capacity and inflate priorities. Budgets align use with company priorities and let idle capacity be borrowed.", "before": "Jobs that could not be preempted held GPUs, teams inflated their priority, and fixed team shares left GPUs idle.", "after": "GPU time is shared across projects by budget, and resumable work can be preempted after a guaranteed run time.", "learn": "Budget GPU time per team. Pair fair share with checkpointing, preemption rules and visible usage data.", "cloudPrimitives": [ "Compute", "Storage" ], "sources": [ { "name": "Hugging Face - Impactful scheduling for GPU clusters (Ai2)", "url": "https://huggingface.co/blog/allenai/impactful-scheduling" } ], "architectureImplication": "Shared GPU fleets can use budget-weighted fair share with preemption. Test checkpointing and fragmentation before rollout.", "tradeoffs": [ "Reliability", "Latency", "Cost", "Control" ], "vendors": [ "Other" ], "decisions": [ "cost-and-routing", "deployment" ] }, { "id": "2026-10-07-claude-haiku-55-copilot", "date": "2026-10-07", "layer": "Model", "kind": "signal", "changeType": "model distribution", "impact": "Packaging / distribution", "title": "GitHub brings Claude Haiku 5.5 to all Copilot plans and surfaces", "whatChanged": "GitHub made Claude Haiku 5.5 generally available in Copilot for individual, Business and Enterprise plans, across IDEs, Copilot CLI, cloud agent and github.com. GitHub positions it for fast, high-volume work such as subagents and quick edits. Usage is billed at provider list price.", "whyCare": "Teams get a small, fast model for high-volume subtasks inside their existing coding tools.", "before": "Teams had fewer integrated options for a fast model for subagents and routine coding steps.", "after": "Organizations can enable Haiku 5.5 across Copilot under their model policies.", "learn": "Test it on your coding and subagent tasks. Compare success rate, latency, retries and cost per finished task.", "cloudPrimitives": [ "Compute", "Security" ], "sources": [ { "name": "GitHub Changelog - Claude Haiku 5.5 in GitHub Copilot", "url": "https://github.blog/changelog/2026-10-07-claude-haiku-5-5-in-github-copilot" } ], "architectureImplication": "Add Haiku 5.5 as a routing option for short, high-volume tasks. Measure cost per successful task.", "tradeoffs": [ "Accuracy", "Latency", "Cost", "Reliability", "Control" ], "vendors": [ "GitHub", "Anthropic" ], "decisions": [ "cost-and-routing", "model-selection" ] }, { "id": "2026-10-08-google-gemini-work-agent", "date": "2026-10-08", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "enterprise workplace agent", "impact": "Packaging / distribution", "title": "Google announces a Gemini agent with its own identity for workplace tasks", "whatChanged": "At Gemini at Work 2026, Google announced a Gemini agent that works across Workspace and connected systems. Team agents get their own accounts and identities, and their access follows existing sharing permissions. Google also described audit trails, sandboxed execution, model routing and spend caps.", "whyCare": "Agent identity, permissions, audit and cost limits now come packaged in a workplace suite.", "before": "Enterprises combined chat assistants with separate automations, and handled identity, audit and cost controls in different services.", "after": "Team agents have their own identities, inherit access from shared resources and run under central controls.", "learn": "Test agent-owned identity and user-delegated access on one bounded workflow. Check audit trails and spend limits.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Google Cloud - Welcome to Gemini at Work 2026", "url": "https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026" }, { "name": "Google - Google Cloud introduces the Gemini agent", "url": "https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/gemini-at-work/" } ], "architectureImplication": "Compare suite-native agents with custom agent stacks on identity, permissions, audit and lock-in. Check availability first.", "tradeoffs": [ "Reliability", "Cost", "Security", "Control" ], "candidate": { "status": "watching", "flaggedOn": "2026-10-17", "tests": { "datedEvent": "Google announced it at Gemini at Work on 8 October 2026.", "primarySource": "Google Cloud blog and Google blog posts of 8 October 2026.", "designChoice": "Gives agents their own identity and permissions, separate from the user who delegates to them.", "firstOrDefining": "Qualifies if other major vendors also give agents their own identities, or if this becomes generally available and widely used." } }, "vendors": [ "Google", "Google Cloud" ], "decisions": [ "identity-and-permissions" ], "short": "Agent identity" }, { "id": "2026-10-09-copilot-local-sandbox-ga", "date": "2026-10-09", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "agent sandbox availability", "impact": "Packaging / distribution", "title": "GitHub makes local sandboxing generally available for Copilot coding agents", "whatChanged": "GitHub made local sandboxing generally available at no extra cost in Copilot CLI, the Copilot app and VS Code Agent Host sessions. It limits agent commands' access to files, networks and credentials. Enterprise settings can require it.", "whyCare": "Coding agents run commands with the developer's privileges. A built-in sandbox makes limiting that access simpler.", "before": "Copilot CLI commands ran with the user's full access unless teams added other controls.", "after": "Organizations can turn on OS-level sandboxing across several Copilot surfaces and enforce it centrally.", "learn": "Test the sandbox's file, network and credential limits on each operating system. Use cloud sandboxes where you need stronger isolation.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "GitHub Changelog - GitHub Copilot weekly releases, October 5", "url": "https://github.blog/changelog/2026-10-09-github-copilot-weekly-releases-october-5" }, { "name": "GitHub Docs - About cloud and local sandboxes for GitHub Copilot", "url": "https://docs.github.com/en/copilot/concepts/security-governance-and-network-settings/about-cloud-and-local-sandboxes" } ], "architectureImplication": "Compare local sandboxing with containers, VMs and GitHub's cloud sandbox. Match the isolation level to your threat model.", "tradeoffs": [ "Security", "Control", "Reliability", "Cost" ], "vendors": [ "GitHub" ], "decisions": [ "security-and-guardrails" ] }, { "id": "2026-10-10-copilot-jetbrains-agent-controls", "date": "2026-10-10", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "agent configuration controls", "impact": "Incremental", "title": "GitHub adds managed model defaults and MCP startup controls to Copilot for JetBrains", "whatChanged": "GitHub added enterprise settings to choose the default Copilot agent model for new JetBrains conversations. Users can still pick another model. A new setting can stop configured MCP servers from starting automatically.", "whyCare": "Administrators get a consistent starting model, and developers control when tools become active. Neither setting restricts which models or MCP tools are allowed.", "before": "JetBrains users chose models with no managed default, and MCP servers started automatically.", "after": "Administrators set the default model, and users can turn off automatic MCP startup.", "learn": "Use model allowlists or a gateway where you must restrict models. Keep MCP permissions and least privilege separate from startup settings.", "cloudPrimitives": [ "Compute", "Networking", "Security" ], "sources": [ { "name": "GitHub Changelog - New controls and chat improvements in Copilot for JetBrains", "url": "https://github.blog/changelog/2026-10-10-new-controls-and-chat-improvements-in-copilot-for-jetbrains" } ], "architectureImplication": "Treat the default model as a preference. Enforce model policy and MCP permissions through separate controls.", "tradeoffs": [ "Control", "Security", "Compatibility" ], "vendors": [ "GitHub" ], "decisions": [ "model-selection", "security-and-guardrails" ] }, { "id": "2026-10-12-agentcore-stm-pricing", "date": "2026-10-12", "layer": "Production & Business", "kind": "signal", "changeType": "agent memory pricing", "impact": "Incremental", "title": "AWS AgentCore now charges for short-term memory by data volume", "whatChanged": "AWS replaced the flat per-event price for AgentCore short-term memory with separate charges: $1.00 per GB ingested, $0.20 per GB retrieved and $0.10 per GB-month stored. Ingestion and retrieval bill at least 12 KB and at most 64 KB per event. Long-term memory pricing stays the same.", "whyCare": "Agent memory cost now depends on payload size, retrieval volume and retention, so event counts alone no longer predict it.", "before": "AWS charged a flat rate per event, whatever the amount of data.", "after": "AWS charges for data ingested, retrieved and stored, with minimum and maximum billable sizes per event.", "learn": "Recalculate memory costs with your real payload sizes, retrieval rates and retention times. Compare total cost with other memory options.", "cloudPrimitives": [ "Storage", "Compute" ], "sources": [ { "name": "AWS - Consumption-based pricing for short-term memory in Amazon Bedrock AgentCore", "url": "https://aws.amazon.com/about-aws/whats-new/2026/10/agentcore-memory-consumption-based-pricing" }, { "name": "AWS - Amazon Bedrock AgentCore pricing", "url": "https://aws.amazon.com/bedrock/agentcore/pricing/" } ], "architectureImplication": "Size memory payloads, retrieval patterns and retention times on purpose, because each now drives cost.", "tradeoffs": [ "Cost", "Control", "Complexity" ], "vendors": [ "AWS" ], "decisions": [ "cost-and-routing", "state-and-memory" ], "action": { "type": "pricing", "what": "Recalculate agent memory costs with payload size, retrieval volume and retention." } }, { "id": "2026-10-12-copilot-telemetry-controls", "date": "2026-10-12", "layer": "Production & Business", "kind": "signal", "changeType": "AI telemetry and governance", "impact": "Incremental", "title": "GitHub adds identity attribution and per-category telemetry controls to Copilot", "whatChanged": "GitHub added an opt-in setting that records the Copilot account, OS account and host name in OpenTelemetry data. Enterprise settings separately control capture of prompts, responses, tool arguments, tool output and policy details. All capture is off by default.", "whyCare": "Enterprises can link agent activity to users and devices without collecting conversation content.", "before": "Telemetry used a pseudonymous user ID, and one setting controlled all content capture.", "after": "Administrators control identity capture and five content types separately, all off by default.", "learn": "Collect the minimum data you need. Content capture can include source code, tool data and local paths.", "cloudPrimitives": [ "Security", "Storage" ], "sources": [ { "name": "GitHub Changelog - Client telemetry user attribution and configurable capture", "url": "https://github.blog/changelog/2026-10-12-client-telemetry-user-attribution-and-configurable-capture" }, { "name": "GitHub Docs - Enterprise managed settings", "url": "https://docs.github.com/copilot/reference/enterprise-administrators/enterprise-managed-settings" } ], "architectureImplication": "Keep identity and content telemetry as separate policy decisions, set centrally.", "tradeoffs": [ "Privacy", "Auditability", "Control" ], "vendors": [ "GitHub" ], "decisions": [ "observability-and-evals", "identity-and-permissions" ] }, { "id": "2026-10-12-lambda-web-functions", "date": "2026-10-12", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "serverless HTTP compute", "impact": "Packaging / distribution", "title": "AWS launches Lambda Web Functions for HTTP servers and streaming AI endpoints", "whatChanged": "AWS introduced Lambda Web Functions, which run standard Node.js HTTP servers, including streaming inference endpoints. One warm environment serves up to 128 concurrent requests. Compute is billed on active CPU time and memory for the environment's lifetime. It launches on Node.js 24, Arm64 and five Regions.", "whyCare": "Teams can deploy an HTTP server without rewriting it as an event handler, and pay less for CPU while waiting on I/O.", "before": "Lambda required an event-handler model, so HTTP apps needed adapters or separate hosting.", "after": "A Node.js HTTP server runs on Lambda with concurrency, native streaming and active-CPU billing.", "learn": "Compare it with event-driven functions and containers on concurrency, memory cost, runtime and Region limits.", "cloudPrimitives": [ "Compute", "Networking", "Security" ], "sources": [ { "name": "AWS News Blog - Announcing AWS Lambda Web Functions", "url": "https://aws.amazon.com/blogs/aws/announcing-aws-lambda-web-functions-for-fast-consistent-web-app-performance-anywhere/" }, { "name": "AWS What's New - AWS Lambda announces Lambda Web Functions", "url": "https://aws.amazon.com/about-aws/whats-new/2026/10/aws-lambda-web-functions/" } ], "architectureImplication": "Consider HTTP-native serverless hosting for streaming inference APIs with long I/O waits.", "tradeoffs": [ "Latency", "Cost", "Control", "Portability" ], "vendors": [ "AWS" ], "decisions": [ "deployment" ] }, { "id": "2026-10-13-gemini-4-argon", "date": "2026-10-13", "layer": "Model", "kind": "signal", "changeType": "frontier multimodal model", "impact": "Step-change", "title": "Google releases Gemini 4 Argon, a frontier model for complex reasoning", "whatChanged": "Google announced Gemini 4 Argon as its newest frontier model and made it available to developers and enterprises. Google describes visual and spatial reasoning, physical understanding and code generation. It reports an 87% score on its own Embodied Reasoning benchmark and an internal video-to-simulation-to-robot demonstration.", "whyCare": "Teams building vision-driven automation or robotics get a new model to test. The benchmark and robot demonstration are Google's own results.", "before": "Teams had no access to Gemini 4 Argon or its stated physical-reasoning abilities.", "after": "Developers and enterprises can test Gemini 4 Argon for complex reasoning and embodied-AI work.", "learn": "Test it on your own video, spatial and action tasks. Compare quality, latency, cost and controls with the models you use now.", "cloudPrimitives": [ "Compute", "Networking" ], "sources": [ { "name": "Google AI Blog - Gemini 4 Argon: Built to solve your hardest problems", "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/" } ], "architectureImplication": "Add Gemini 4 Argon as a candidate for video and spatial work. Measure quality and serving cost before you depend on it.", "tradeoffs": [ "Accuracy", "Latency", "Cost", "Control" ], "vendors": [ "Google" ], "decisions": [ "model-selection" ] }, { "id": "2026-10-13-agent-registry-custom-metadata", "date": "2026-10-13", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "agent registry governance", "impact": "Incremental", "title": "AWS Agent Registry now checks agent metadata against a schema", "whatChanged": "AWS added custom metadata schemas to Agent Registry for agents, MCP servers, skills, gateways and other records. Teams can require fields such as owner, data classification or review status. The registry checks values when records change and lets you filter by them.", "whyCare": "Teams with many agents and tools can record owners and review status in the catalog itself, in a consistent form.", "before": "Registry records had names and descriptions, but no required fields that the registry checked.", "after": "Administrators define schemas, enforce required fields and allowed values, and filter by them.", "learn": "Start with a small shared schema for owner, classification and lifecycle status. Keep access control and approvals in their own systems.", "cloudPrimitives": [ "Storage", "Security", "Networking" ], "sources": [ { "name": "AWS What's New - AWS Agent Registry now supports custom metadata for registry records", "url": "https://aws.amazon.com/about-aws/whats-new/2026/10/aws-agent-registry-custom-metadata-registry-records" }, { "name": "AWS Documentation - Define a custom metadata schema", "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/registry-custom-metadata-schema.html" } ], "architectureImplication": "Use the registry to find and govern agent resources. Keep authorization and policy checks in the runtime and connected services.", "tradeoffs": [ "Control", "Security", "Complexity", "Portability" ], "vendors": [ "AWS" ], "decisions": [ "identity-and-permissions" ] }, { "id": "2026-10-13-langchain-deep-agents-deploy", "date": "2026-10-13", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "open agent deployment", "impact": "Packaging / distribution", "title": "LangChain offers a self-hostable way to deploy long-running agents on any model", "whatChanged": "LangChain launched Deep Agents deploy in beta. It bundles its open-source agent harness with a scalable LangSmith Deployment server, with endpoints for MCP, A2A, human approvals and memory. It works with several model and sandbox providers, and LangChain says teams can self-host it to keep control of agent memory.", "whyCare": "Teams get a middle path between building their own agent runtime and adopting one vendor's managed agent service.", "before": "Teams either ran their own orchestration, memory and sandbox around an open framework, or used a vendor's managed runtime.", "after": "One command deploys an open, model-agnostic harness, managed or self-hosted, with standard agent endpoints.", "learn": "Compare it with self-hosting and with managed services such as Bedrock AgentCore. Before you store production state in it, test memory export, sandbox isolation, approvals and how reliable the beta is.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "LangChain - Deep Agents Deploy: an open alternative to Claude Managed Agents", "url": "https://www.langchain.com/blog/deep-agents-deploy-an-open-alternative-to-claude-managed-agents" }, { "name": "LangChain documentation - Managed Deep Agents", "url": "https://docs.langchain.com/langsmith/python/managed-deep-agents-overview" } ], "architectureImplication": "Treat the harness, runtime, memory and sandbox as separate lock-in decisions. A model-agnostic harness alone does not remove lock-in.", "tradeoffs": [ "Portability", "Control", "Security", "Complexity", "Reliability" ], "vendors": [ "LangChain" ], "decisions": [ "managed-vs-custom-runtime" ] }, { "id": "2026-10-14-github-hydrafusion", "date": "2026-10-14", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "multi-model orchestration", "impact": "Packaging / distribution", "title": "GitHub Copilot HydraFusion drafts and reviews with several models per task", "whatChanged": "GitHub updated HydraFusion, its research-preview Copilot orchestrator. Cascade judges a draft and asks a stronger model for a repair only when needed. Critique runs models in parallel to inspect the work. Each model call is billed at its standard rate.", "whyCare": "Teams get draft-and-review patterns without building their own harness, though multi-pass work costs more credits and time.", "before": "Teams picked one model, or built escalation and review workflows into their own agent harness.", "after": "Copilot can pick one model, a judged cascade or parallel critique for each task.", "learn": "Compare it with single-model use on real coding tasks. Track accepted changes, regressions, latency and credits. Keep it out of critical workflows while in preview.", "cloudPrimitives": [ "Compute", "Networking", "Security" ], "sources": [ { "name": "GitHub Changelog - HydraFusion updates and new Max option", "url": "https://github.blog/changelog/2026-10-14-hydrafusion-updates-and-new-max-option" }, { "name": "GitHub documentation - Using HydraFusion", "url": "https://docs.github.com/en/early-access/copilot/hydrafusion" } ], "architectureImplication": "Choose orchestration per workload based on quality and cost. The preview has no SLA.", "tradeoffs": [ "Accuracy", "Latency", "Cost", "Control", "Reliability" ], "vendors": [ "GitHub" ], "decisions": [ "cost-and-routing" ] }, { "id": "2026-10-15-sagemaker-muse-spark-open", "date": "2026-10-15", "layer": "Model", "kind": "signal", "changeType": "managed open-weight model deployment", "impact": "Packaging / distribution", "title": "AWS adds Meta Muse Spark Open to SageMaker JumpStart for in-account deployment", "whatChanged": "AWS made Meta's Muse Spark Open available in SageMaker JumpStart in BF16 and NVFP4 variants. The 816B-parameter model has 42B active parameters, a 1M-token context window and text, image, document, audio and video input. It launches in two US East Regions.", "whyCare": "Teams can test a large open-weight model on private code and media without sending data to a third-party API.", "before": "Teams needed another deployment route to run Muse Spark Open on AWS.", "after": "Teams can deploy it as a SageMaker endpoint in their own account with VPC, KMS encryption and network isolation.", "learn": "Benchmark both variants on your tasks. Check the license, quotas and cost, since the endpoint bills on large GPU instances while active.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "AWS Machine Learning Blog - Meta's Muse Spark Open now available in Amazon SageMaker JumpStart", "url": "https://aws.amazon.com/blogs/machine-learning/metas-muse-spark-open-now-available-in-amazon-sagemaker-jumpstart/" } ], "architectureImplication": "Weigh open-weight control against hosted APIs. Size for the full model memory and enforce license, network and cost controls.", "tradeoffs": [ "Cost", "Control", "Security", "Latency", "Complexity" ], "vendors": [ "AWS", "Meta", "Open models" ], "decisions": [ "hosted-vs-open", "deployment" ] }, { "id": "2026-10-15-github-copilot-model-deprecations", "date": "2026-10-15", "layer": "Production & Business", "kind": "signal", "changeType": "model lifecycle and migration", "impact": "Incremental", "title": "GitHub retires three Copilot models on 16 November 2026", "whatChanged": "GitHub will deprecate Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra across Copilot on 16 November 2026. It recommends Claude Opus 5.5, GPT-6.1 Sol or GPT-6 Luna as replacements, enabled automatically under default model policies.", "whyCare": "Teams using these models have a fixed migration deadline and need to test the replacements first.", "before": "The three models were available across supported Copilot experiences.", "after": "They disappear from most Copilot experiences after 16 November. Opus 4.8 stays for annual individual plans.", "learn": "List where you use these models. Test the replacements on real tasks and check your model policies before the deadline.", "cloudPrimitives": [ "Compute", "Security" ], "sources": [ { "name": "GitHub Changelog - Upcoming deprecation of selected GitHub Copilot models in mid-November", "url": "https://github.blog/changelog/2026-10-15-upcoming-deprecation-of-selected-github-copilot-models-in-mid-november" }, { "name": "GitHub Docs - Supported AI models in GitHub Copilot", "url": "https://docs.github.com/copilot/reference/ai-models/supported-models" } ], "architectureImplication": "Treat hosted models as versioned dependencies. Keep tests ready so you can switch models before a vendor retires one.", "tradeoffs": [ "Compatibility", "Accuracy", "Control", "Reliability" ], "vendors": [ "GitHub", "Anthropic", "OpenAI" ], "decisions": [ "model-selection" ], "action": { "type": "retirement", "by": "2026-11-16", "what": "Test the replacement models and update your Copilot model policies." } }, { "id": "2026-10-16-github-open-source-license-compliance-ga", "date": "2026-10-16", "layer": "Production & Business", "kind": "signal", "changeType": "agent supply-chain policy enforcement", "impact": "Incremental", "title": "GitHub enforces license policy on new dependencies, including those added by agents", "whatChanged": "GitHub made open source license compliance generally available for Enterprise Cloud with Code Security. Rulesets can block merges until new dependencies meet license policy. A new API lets coding agents check a package before they add it.", "whyCare": "Coding agents add dependencies quickly, so license checks need to happen when they choose a package and again at merge.", "before": "Pull request checks existed, but agents had no generally available API to check a package against policy.", "after": "Agents can check packages through the API, and rulesets enforce policy before merge.", "learn": "Connect your coding agents to the package check. Keep the required merge check as the final gate.", "cloudPrimitives": [ "Compute", "Security" ], "sources": [ { "name": "GitHub Changelog - Open source license compliance generally available", "url": "https://github.blog/changelog/2026-10-16-open-source-license-compliance-generally-available" }, { "name": "GitHub Docs - About open source license compliance", "url": "https://docs.github.com/enterprise-cloud@latest/code-security/concepts/supply-chain-security/open-source-license-compliance" } ], "architectureImplication": "Check license policy at package selection and enforce it again at merge. Audit exceptions.", "tradeoffs": [ "Security", "Control", "Complexity" ], "vendors": [ "GitHub" ], "decisions": [ "security-and-guardrails", "compliance" ] }, { "id": "2026-10-16-reflection-beam-open-model-aws", "date": "2026-10-16", "layer": "Model", "kind": "signal", "changeType": "open-weight model and deployment availability", "impact": "Packaging / distribution", "title": "Reflection releases Beam, an open coding model you can run on AWS", "whatChanged": "Reflection released Beam 1.0, an open-weight coding model with 501B total and 23B active parameters and a reported 1M-token context, under Apache 2.0. AWS offers it through SageMaker JumpStart and HyperPod in two US East Regions.", "whyCare": "You can now run a coding model for agents inside your own AWS account, under a license that allows commercial use.", "before": "Teams had to assemble serving for other open coding models themselves, and Beam was not yet available.", "after": "Teams can deploy Beam through SageMaker JumpStart or HyperPod, or run the weights themselves.", "learn": "Test Beam on your own agent tasks before you trust its benchmarks. Plan memory for all 501B parameters. Check which regions offer it.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Hugging Face Blog - Welcome Beam: a 501B open model you can own, customize, and scale", "url": "https://huggingface.co/blog/reflection-beam" }, { "name": "AWS Machine Learning Blog - Introducing Reflection Beam, an open-weight model for agentic coding on AWS", "url": "https://aws.amazon.com/blogs/machine-learning/introducing-reflection-beam-an-open-weight-model-for-agentic-coding-on-aws/" }, { "name": "AWS What's New - Beam model by Reflection is now available on Amazon SageMaker JumpStart and AWS Trainium", "url": "https://aws.amazon.com/about-aws/whats-new/2026/01/reflection-beam-on-sagemaker-jumpstart-and-aws-trainium/" } ], "architectureImplication": "Compare self-hosting Beam with hosted coding models on quality, serving footprint and cost.", "tradeoffs": [ "Cost", "Control", "Complexity", "Latency" ], "vendors": [ "AWS", "Open models", "Other" ], "decisions": [ "hosted-vs-open", "deployment" ] }, { "id": "2023-05-11-claude-100k-context", "date": "2023-05-11", "layer": "Data & Knowledge", "kind": "milestone", "changeType": "long context", "impact": "Step-change", "title": "Anthropic expands Claude's context window to 100K tokens", "whatChanged": "Anthropic raised Claude's context window from 9K to 100K tokens, about 75,000 words.", "whyCare": "Teams could put whole documents into one prompt for the first time at this scale.", "before": "Most models held a few thousand tokens, so long documents had to be split or summarised first.", "after": "A single request could hold a long report, contract or codebase section.", "learn": "Compare long context with retrieval on your own documents. Measure cost and accuracy for each.", "cloudPrimitives": [ "Compute", "Storage" ], "sources": [ { "name": "Anthropic - Introducing 100K context windows", "url": "https://www.anthropic.com/news/100k-context-windows" } ], "architectureImplication": "Long context became a design option next to retrieval. Choose per use case based on cost, latency and accuracy.", "tradeoffs": [ "Cost", "Latency", "Accuracy" ], "verified": "2026-10-17", "vendors": [ "Anthropic" ], "decisions": [ "rag-vs-long-context" ], "short": "100K context" }, { "id": "2023-07-18-llama-2", "date": "2023-07-18", "layer": "Model", "kind": "milestone", "changeType": "open-weight model", "impact": "Step-change", "title": "Meta releases Llama 2 free for research and commercial use", "whatChanged": "Meta released Llama 2 with model weights that companies could download and use commercially, together with Microsoft as a launch partner.", "whyCare": "Companies could run a capable model on their own infrastructure under a license that allowed business use.", "before": "Strong models were mostly available only through hosted APIs, and open models were limited or research-only.", "after": "Teams could self-host, fine-tune and control a capable model.", "learn": "Check the license terms for your use case. Plan for hosting, updates and safety controls you now own.", "cloudPrimitives": [ "Compute", "Storage", "Security" ], "sources": [ { "name": "Meta - Meta and Microsoft Introduce the Next Generation of Llama", "url": "https://about.fb.com/news/2023/07/llama-2/" } ], "architectureImplication": "Self-hosted models became a real option for data control and cost. Teams take on serving, scaling and model updates.", "tradeoffs": [ "Control", "Cost", "Complexity" ], "verified": "2026-10-17", "vendors": [ "Meta", "Open models" ], "decisions": [ "hosted-vs-open" ], "short": "Llama 2" }, { "id": "2023-08-02-owasp-llm-top-10", "date": "2023-08-02", "layer": "Production & Business", "kind": "milestone", "changeType": "security standard", "impact": "Step-change", "title": "OWASP publishes the first Top 10 risks for LLM applications", "whatChanged": "OWASP published version 1.0 of its Top 10 for Large Language Model Applications. It lists the main risks, such as prompt injection and insecure output handling.", "whyCare": "Security teams got a shared list to review LLM applications against.", "before": "Teams had no common reference for LLM-specific security risks.", "after": "Reviews, threat models and vendor questionnaires could point to one public list.", "learn": "Review your application against the current version of the list. Add the relevant risks to your threat model.", "cloudPrimitives": [ "Security" ], "sources": [ { "name": "OWASP - Top 10 for Large Language Model Applications", "url": "https://owasp.org/www-project-top-10-for-large-language-model-applications/" }, { "name": "SD Times - OWASP releases Top 10 list for LLMs", "url": "https://sdtimes.com/security/owasp-releases-top-10-list-for-llms/" } ], "architectureImplication": "Treat model input and output as untrusted. Add controls for prompt injection, output handling and excessive agent permissions.", "tradeoffs": [ "Security", "Control" ], "verified": "2026-10-17", "verificationNote": "The OWASP page shows no release date. Date from SD Times coverage published 2 August 2023.", "vendors": [ "Open standards" ], "decisions": [ "security-and-guardrails" ], "short": "OWASP LLM Top 10" }, { "id": "2023-11-01-microsoft-365-copilot-ga", "date": "2023-11-01", "layer": "Application & Workflow", "kind": "milestone", "changeType": "workplace AI assistant", "impact": "Step-change", "title": "Microsoft makes Microsoft 365 Copilot generally available to enterprises", "whatChanged": "Microsoft made Copilot for Microsoft 365 generally available to enterprise customers on 1 November 2023, inside Word, Excel, PowerPoint, Outlook and Teams.", "whyCare": "GenAI moved into the office software most companies already use, with access tied to existing company data and permissions.", "before": "Most GenAI use at work happened in separate chat tools.", "after": "Employees could use GenAI inside their daily applications, grounded in company files and email.", "learn": "Check how your data permissions apply when an assistant can read across files and email. Fix oversharing before rollout.", "cloudPrimitives": [ "Storage", "Security" ], "sources": [ { "name": "Microsoft - Announcing Microsoft 365 Copilot general availability and Microsoft 365 Chat", "url": "https://www.microsoft.com/en-us/copilot/blog/2023/09/21/announcing-microsoft-365-copilot-general-availability-and-microsoft-365-chat/" } ], "architectureImplication": "Existing access controls became the boundary for what AI can see. Data governance became a prerequisite for rollout.", "tradeoffs": [ "Security", "Control" ], "verified": "2026-10-17", "vendors": [ "Microsoft" ], "decisions": [ "build-vs-buy", "identity-and-permissions" ], "short": "Microsoft 365 Copilot" }, { "id": "2023-11-06-openai-devday-assistants", "date": "2023-11-06", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "managed assistant runtime", "impact": "Step-change", "title": "OpenAI launches the Assistants API and GPT-4 Turbo with 128K context", "whatChanged": "At DevDay, OpenAI announced GPT-4 Turbo with a 128K context window, JSON mode, and the Assistants API with threads, retrieval and code execution.", "whyCare": "Developers could build assistants that keep state and use tools without running that infrastructure themselves.", "before": "Developers stored conversation state, retrieval and tool loops in their own code.", "after": "The platform could hold threads, files and tool calls for them.", "learn": "Decide which state you want the vendor to hold. Check how to export it if you change vendors.", "cloudPrimitives": [ "Compute", "Storage", "Networking" ], "sources": [ { "name": "OpenAI - New models and developer products announced at DevDay", "url": "https://openai.com/index/new-models-and-developer-products-announced-at-devday/" } ], "architectureImplication": "State and tool orchestration could move into the vendor platform. This trades build effort for portability.", "tradeoffs": [ "Control", "Portability", "Complexity" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "managed-vs-custom-runtime", "state-and-memory" ], "short": "Assistants API" }, { "id": "2024-08-01-eu-ai-act-in-force", "date": "2024-08-01", "layer": "Production & Business", "kind": "milestone", "changeType": "regulation", "impact": "Step-change", "title": "EU AI Act enters into force", "whatChanged": "The EU AI Act entered into force on 1 August 2024. Its rules apply in stages over the following years.", "whyCare": "Teams that build or use AI in the EU now have legal duties that depend on the risk level of each use case.", "before": "No broad law set rules for how AI systems are built and used in the EU.", "after": "Providers and deployers must classify use cases by risk and meet the duties for each level.", "learn": "List your AI use cases and classify them by risk. Track which obligations apply on which date.", "cloudPrimitives": [ "Security" ], "sources": [ { "name": "European Commission - AI Act enters into force", "url": "https://commission.europa.eu/news-and-media/news/ai-act-enters-force-2024-08-01_en" } ], "architectureImplication": "Plan for documentation, logging, human oversight and transparency features as part of the design.", "tradeoffs": [ "Control", "Auditability", "Cost" ], "verified": "2026-10-17", "vendors": [ "EU regulation" ], "decisions": [ "compliance" ], "short": "EU AI Act" }, { "id": "2024-08-06-openai-structured-outputs", "date": "2024-08-06", "layer": "Data & Knowledge", "kind": "milestone", "changeType": "structured output", "impact": "Step-change", "title": "OpenAI makes model output follow a developer's JSON schema", "whatChanged": "OpenAI introduced Structured Outputs in the API. Model outputs now reliably follow a JSON Schema the developer supplies.", "whyCare": "Systems that pass model output to other code can rely on its shape.", "before": "Developers asked for JSON in the prompt, then validated and retried when the output broke.", "after": "The API guarantees the output matches the schema.", "learn": "Define schemas for every model output that other code reads. Remove retry logic that only fixed format errors.", "cloudPrimitives": [ "Compute" ], "sources": [ { "name": "OpenAI - Introducing Structured Outputs in the API", "url": "https://openai.com/index/introducing-structured-outputs-in-the-api/" } ], "architectureImplication": "Model output can feed other systems directly. Validation still matters for content, though format errors drop away.", "tradeoffs": [ "Reliability", "Accuracy" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "integration-standards" ], "short": "Structured outputs" }, { "id": "2024-08-14-anthropic-prompt-caching", "date": "2024-08-14", "layer": "Data & Knowledge", "kind": "milestone", "changeType": "prompt caching", "impact": "Step-change", "title": "Anthropic introduces prompt caching for long, repeated context", "whatChanged": "Anthropic launched prompt caching for Claude. Reusing a cached prompt prefix cuts cost by up to 90% and latency by up to 85% for long prompts.", "whyCare": "Long system prompts, documents and tool definitions became cheap to resend on every call.", "before": "Every request paid full price for the whole prompt, even when most of it repeated.", "after": "Repeated context costs a fraction of the normal input price.", "learn": "Put stable content at the start of the prompt and changing content at the end. Measure your cache hit rate.", "cloudPrimitives": [ "Compute", "Storage" ], "sources": [ { "name": "Anthropic - Prompt caching with Claude", "url": "https://claude.com/blog/prompt-caching" }, { "name": "SD Times - Anthropic introduces prompt caching", "url": "https://sdtimes.com/ai/anthropic-introduces-prompt-caching-to-reduce-latency-and-costs" } ], "architectureImplication": "Prompt structure became a cost lever. Agents and long-context designs got much cheaper to run.", "tradeoffs": [ "Cost", "Latency" ], "verified": "2026-10-17", "verificationNote": "Anthropic's page now shows a later republish date. Launch date confirmed by news coverage from 14 August 2024.", "vendors": [ "Anthropic" ], "decisions": [ "cost-and-routing", "rag-vs-long-context" ], "short": "Prompt caching" }, { "id": "2025-02-24-claude-code-preview", "date": "2025-02-24", "layer": "Application & Workflow", "kind": "milestone", "changeType": "coding agent", "impact": "Step-change", "title": "Anthropic launches Claude Code, an agentic coding tool for the terminal", "whatChanged": "Anthropic released Claude Code in limited research preview with Claude 3.7 Sonnet. Developers can hand off engineering tasks from their terminal.", "whyCare": "Coding assistants started to act on a codebase, editing files and running commands.", "before": "Coding assistants mostly suggested code inside an editor for a developer to accept.", "after": "An agent could carry out multi-step engineering tasks under a developer's direction.", "learn": "Set clear permissions for what a coding agent may run. Review its changes like any other contributor's.", "cloudPrimitives": [ "Compute", "Security" ], "sources": [ { "name": "Anthropic - Claude 3.7 Sonnet and Claude Code", "url": "https://www.anthropic.com/news/claude-3-7-sonnet" } ], "architectureImplication": "Developer machines and repositories became places where agents act. Permissions and review gates matter.", "tradeoffs": [ "Security", "Control", "Reliability" ], "verified": "2026-10-17", "vendors": [ "Anthropic" ], "decisions": [ "human-approval", "security-and-guardrails" ], "short": "Claude Code" }, { "id": "2025-05-16-openai-codex-cloud-agent", "date": "2025-05-16", "layer": "Application & Workflow", "kind": "milestone", "changeType": "cloud coding agent", "impact": "Step-change", "title": "OpenAI launches Codex, a cloud coding agent that works on tasks in parallel", "whatChanged": "OpenAI introduced Codex as a cloud-based software engineering agent that can work on many tasks in parallel.", "whyCare": "Coding agents moved off the developer's machine and into managed cloud sandboxes.", "before": "Coding agents ran one task at a time on a developer's machine.", "after": "Teams could queue several coding tasks to agents running in the cloud.", "learn": "Decide which repositories and secrets a cloud agent may access. Track the cost and quality of agent work.", "cloudPrimitives": [ "Compute", "Security", "Networking" ], "sources": [ { "name": "OpenAI - Introducing Codex", "url": "https://openai.com/index/introducing-codex/" } ], "architectureImplication": "Agent work became something you can run in parallel and in the cloud. Sandboxes, secrets and review queues became part of the design.", "tradeoffs": [ "Security", "Cost", "Control" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "managed-vs-custom-runtime", "security-and-guardrails" ], "short": "Codex" }, { "id": "2025-08-02-eu-ai-act-gpai-obligations", "date": "2025-08-02", "layer": "Production & Business", "kind": "milestone", "changeType": "regulation", "impact": "Step-change", "title": "EU AI Act rules for general-purpose AI models start to apply", "whatChanged": "From 2 August 2025, providers of general-purpose AI models in the EU must meet transparency and copyright duties under the AI Act.", "whyCare": "Model providers must now document their models, which gives teams more information when they choose one.", "before": "Model providers had no legal duty in the EU to document training data or model capabilities.", "after": "Providers must publish documentation and a summary of training content.", "learn": "Ask model vendors for their AI Act documentation. Use it in your model selection and risk records.", "cloudPrimitives": [ "Security" ], "sources": [ { "name": "European Commission - General-purpose AI obligations under the AI Act", "url": "https://digital-strategy.ec.europa.eu/en/factpages/general-purpose-ai-obligations-under-ai-act" } ], "architectureImplication": "Use vendor documentation when you select a model and keep it in your risk records.", "tradeoffs": [ "Control", "Auditability" ], "verified": "2026-10-17", "vendors": [ "EU regulation" ], "decisions": [ "compliance", "model-selection" ], "short": "EU model rules" }, { "id": "2025-08-05-openai-gpt-oss", "date": "2025-08-05", "layer": "Model", "kind": "milestone", "changeType": "open-weight model", "impact": "Step-change", "title": "OpenAI releases gpt-oss, its first open-weight models since 2019", "whatChanged": "OpenAI released gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license.", "whyCare": "A leading model lab now offered open weights, so teams could self-host models from the same vendor they use through an API.", "before": "OpenAI models were available only through hosted services.", "after": "Teams could run OpenAI reasoning models on their own infrastructure.", "learn": "Test the open models against your hosted model on the same tasks. Compare total cost, including serving.", "cloudPrimitives": [ "Compute", "Storage", "Security" ], "sources": [ { "name": "OpenAI - Introducing gpt-oss", "url": "https://openai.com/index/introducing-gpt-oss/" } ], "architectureImplication": "Mixed designs became easier: hosted models for some tasks, self-hosted open weights for sensitive or high-volume ones.", "tradeoffs": [ "Control", "Cost", "Portability" ], "verified": "2026-10-17", "vendors": [ "OpenAI", "Open models" ], "decisions": [ "hosted-vs-open" ], "short": "gpt-oss" }, { "id": "2025-08-07-openai-gpt-5", "date": "2025-08-07", "layer": "Model", "kind": "milestone", "changeType": "frontier model with routing", "impact": "Step-change", "title": "OpenAI releases GPT-5 with a router between fast and deep reasoning", "whatChanged": "OpenAI released GPT-5 as the new default in ChatGPT. A router picks a quick model or a deeper reasoning model for each request.", "whyCare": "Choosing between a fast model and a reasoning model became something the system can decide per request.", "before": "Users and developers picked a reasoning or non-reasoning model themselves.", "after": "One entry point routes each request to the right level of reasoning.", "learn": "Test how routing affects your cost and latency. Pin a model version where you need predictable behaviour.", "cloudPrimitives": [ "Compute" ], "sources": [ { "name": "OpenAI - Introducing GPT-5", "url": "https://openai.com/index/introducing-gpt-5/" } ], "architectureImplication": "Routing between model tiers became a standard design pattern, inside vendor products and in your own systems.", "tradeoffs": [ "Cost", "Latency", "Accuracy" ], "verified": "2026-10-17", "vendors": [ "OpenAI" ], "decisions": [ "cost-and-routing", "model-selection" ], "short": "GPT-5" }, { "id": "2025-09-29-platform-context-management", "date": "2025-09-29", "layer": "Data & Knowledge", "kind": "milestone", "changeType": "context management", "impact": "Step-change", "title": "Anthropic adds context editing and a memory tool for long-running agents", "whatChanged": "Anthropic launched context editing, which clears stale tool results as the context fills, and a memory tool that stores information in files outside the context. OpenAI added compaction in December 2025 and Anthropic a Compaction API in February 2026.", "whyCare": "Agents can now run far longer than one context window because the platform manages what stays in context.", "before": "Developers trimmed, summarised and stored agent context with their own code.", "after": "Major platforms offer context editing, memory and compaction as built-in features.", "learn": "Test how your agent behaves when old context is cleared or compacted. Decide what must go to durable memory.", "cloudPrimitives": [ "Storage", "Compute" ], "sources": [ { "name": "Anthropic - Managing context on the Claude Developer Platform", "url": "https://claude.com/blog/context-management" }, { "name": "OpenAI - API changelog (compaction, 11 December 2025)", "url": "https://developers.openai.com/api/docs/changelog" }, { "name": "Anthropic - Claude Platform release notes (Compaction API, 5 February 2026)", "url": "https://platform.claude.com/docs/en/release-notes/overview" } ], "architectureImplication": "Context became a managed resource. Design what lives in the window, what goes to memory and what gets summarised.", "tradeoffs": [ "Reliability", "Cost", "Accuracy" ], "verified": "2026-10-17", "vendors": [ "Anthropic", "OpenAI" ], "decisions": [ "state-and-memory" ], "short": "Context management" }, { "id": "2025-12-09-agentic-ai-foundation", "date": "2025-12-09", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "open standards governance", "impact": "Step-change", "title": "Linux Foundation forms the Agentic AI Foundation for MCP, goose and AGENTS.md", "whatChanged": "The Linux Foundation announced the Agentic AI Foundation, with the Model Context Protocol, goose and AGENTS.md as founding projects.", "whyCare": "Key agent standards moved under neutral governance, which lowers the risk of building on them.", "before": "MCP and related agent standards were led by single companies.", "after": "A vendor-neutral foundation governs them.", "learn": "Prefer standards with neutral governance for long-lived integrations. Track the foundation's specification releases.", "cloudPrimitives": [ "Networking", "Security" ], "sources": [ { "name": "Linux Foundation - Announces the formation of the Agentic AI Foundation", "url": "https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation" } ], "architectureImplication": "MCP-based integrations became a safer long-term bet for tool and data connections.", "tradeoffs": [ "Portability", "Control" ], "verified": "2026-10-17", "vendors": [ "Open standards" ], "decisions": [ "integration-standards" ], "short": "Agentic AI Foundation" }, { "id": "2025-12-18-agent-skills-open-standard", "date": "2025-12-18", "layer": "Agent & Infrastructure", "kind": "milestone", "changeType": "open standard", "impact": "Incremental", "title": "Anthropic publishes Agent Skills as an open standard", "whatChanged": "Anthropic published the Agent Skills specification as an open standard at agentskills.io. OpenAI added Skills support to its Responses API in February 2026.", "whyCare": "Teams can package instructions, scripts and resources for agents once and use them across vendors.", "before": "Each agent platform had its own way to package reusable know-how.", "after": "A shared format works across at least two major platforms.", "learn": "Package repeated agent procedures as skills. Keep them in version control like code.", "cloudPrimitives": [ "Storage" ], "sources": [ { "name": "Anthropic - Organization skills and directory", "url": "https://claude.com/blog/organization-skills-and-directory" }, { "name": "OpenAI - API changelog (Skills in the Responses API, 10 February 2026)", "url": "https://developers.openai.com/api/docs/changelog" } ], "architectureImplication": "Agent know-how became a portable asset that teams can version, review and share.", "tradeoffs": [ "Portability", "Control" ], "verified": "2026-10-17", "vendors": [ "Anthropic", "OpenAI", "Open standards" ], "decisions": [ "integration-standards" ], "short": "Agent Skills" }, { "id": "2026-08-02-eu-ai-act-transparency", "date": "2026-08-02", "layer": "Production & Business", "kind": "milestone", "changeType": "regulation", "impact": "Step-change", "title": "EU AI Act transparency rules start to apply", "whatChanged": "From 2 August 2026, AI-generated content such as deepfakes must be labelled, and users must be told when they interact with an AI system such as a chatbot.", "whyCare": "Customer-facing AI in the EU now needs disclosure and content labelling built in.", "before": "Disclosure and labelling were voluntary in most products.", "after": "Providers and deployers in the EU must disclose AI interaction and label AI-generated content.", "learn": "Check every customer-facing AI feature for disclosure. Add labels or metadata to generated media.", "cloudPrimitives": [ "Security" ], "sources": [ { "name": "European Commission - Safer and more transparent AI", "url": "https://commission.europa.eu/news-and-media/news/safer-and-more-transparent-ai-2026-08-02_en" } ], "architectureImplication": "Disclosure and content labelling became product requirements for AI features used in the EU.", "tradeoffs": [ "Control", "Auditability" ], "verified": "2026-10-17", "vendors": [ "EU regulation" ], "decisions": [ "compliance" ], "short": "EU transparency rules" }, { "id": "2026-10-15-github-copilot-multi-repo-slack-teams", "date": "2026-10-15", "layer": "Application & Workflow", "kind": "signal", "changeType": "cross-repository coding agent workflow", "impact": "Incremental", "title": "GitHub lets Copilot span repositories in Slack and Teams", "whatChanged": "GitHub says Copilot in Slack and Microsoft Teams now works across several repositories in one conversation. Users can create issues, open pull requests and investigate bugs across them. The update also shares Slack thread context and makes switching repositories easier.", "whyCare": "Teams that coordinate work across services can handle cross-repository tasks in one chat, with repository permissions and pull request review still in place.", "before": "Copilot in Slack and Teams handled one repository per conversation.", "after": "One conversation can span repositories, from investigating a bug to opening a pull request.", "learn": "Test cross-repository tasks in representative Slack and Teams workflows. Review which repositories your users can access and preserve human review before merging agent-generated changes.", "cloudPrimitives": [ "Compute", "Networking", "Security" ], "sources": [ { "name": "GitHub Changelog - Multi-repository support and other improvements for GitHub in Slack and Teams", "url": "https://github.blog/changelog/2026-10-15-copilot-in-slack-and-teams-adds-model-and-repository-updates" } ], "architectureImplication": "Teams can start cross-repository agent tasks in Slack and Teams and carry thread context into code work. Keep repository access scoped, and retain pull request checks and human review.", "tradeoffs": [ "Security", "Control", "Complexity" ], "vendors": [ "GitHub" ], "decisions": [] }, { "id": "2026-10-15-ibm-granite-switch-4-1", "date": "2026-10-15", "layer": "Model", "kind": "signal", "changeType": "multi-task inference", "impact": "Packaging / distribution", "title": "IBM combines chat, decisions, and speech in Granite-Switch-4.1", "whatChanged": "IBM released Granite-Switch-4.1 in 8B and 30B sizes on 15 October. Each model combines a decision adapter, a safety adapter called Guardian, and a speech model for transcription. IBM says the parts use about the same GPU memory as the base model and can share its KV cache.", "whyCare": "You can serve chat, transcription and specialised decisions from one endpoint, with less extra GPU memory. Compare it with separate specialist models first.", "before": "Teams combined task adapters themselves or sent chat, classification and speech to separate models.", "after": "One Granite Switch endpoint handles chat, decisions and transcription. It runs on vLLM through a plugin and in llama.cpp.", "learn": "Compare Granite Switch with separate speech and decision models on accuracy, latency, throughput and GPU memory.", "cloudPrimitives": [ "Compute", "Networking" ], "sources": [ { "name": "IBM Granite on Hugging Face - Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint", "url": "https://huggingface.co/blog/ibm-granite/granite-switch-4-1" }, { "name": "Granite Switch project", "url": "https://github.com/generative-computing/granite-switch" } ], "architectureImplication": "One multi-adapter model can replace several endpoints. Keep per-task tests and fallback routes, since equal quality across tasks is not shown.", "tradeoffs": [ "Cost", "Accuracy", "Latency", "Complexity" ], "vendors": [ "IBM" ], "decisions": [ "model-selection", "deployment" ] }, { "id": "2026-10-20-google-antigravity-enterprise", "date": "2026-10-20", "layer": "Application & Workflow", "kind": "signal", "changeType": "enterprise agent distribution", "impact": "Packaging / distribution", "title": "Google Cloud includes Antigravity in Gemini Enterprise subscriptions", "whatChanged": "Google Cloud made Google Antigravity available in eligible Gemini Enterprise subscriptions on 20 October. Admins can control project spend, pooled quotas, overages, sandboxing, browser and MCP access, and audit logging. IDE extensions bring Antigravity to VS Code, Visual Studio, JetBrains, and Zed; some integrations remain in preview.", "whyCare": "You can give coding teams an agent across their preferred editors while using Gemini Enterprise for identity, spend and governance controls. Check which extensions are generally available before rollout.", "before": "Teams managed Antigravity separately from Gemini Enterprise subscriptions, including its tools and admin controls.", "after": "Eligible Gemini Enterprise subscriptions now include Antigravity, along with pooled quotas, project spend caps, usage metrics, audit logs and policy controls across supported surfaces.", "learn": "Check license eligibility and preview status for each IDE. Set spend caps, sandbox and MCP policies, and audit settings before you roll it out.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Google Cloud - Expanding Google Antigravity for enterprise customers", "url": "https://cloud.google.com/blog/products/ai-machine-learning/expanding-google-antigravity-for-enterprise-customers" }, { "name": "Google Antigravity - Antigravity IDE Extensions", "url": "https://antigravity.google/blog/antigravity-ide-extensions" } ], "architectureImplication": "Google Cloud now offers Antigravity as a managed coding-agent option within Gemini Enterprise. Compare its license scope, IDE coverage, identity controls, audit collection and quota behavior with your stack.", "tradeoffs": [ "Cost", "Control", "Portability", "Security" ], "vendors": [ "Google", "Google Cloud" ], "decisions": [ "build-vs-buy", "identity-and-permissions", "security-and-guardrails" ] }, { "id": "2026-10-20-torchneuron", "date": "2026-10-20", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "native pytorch backend", "impact": "Packaging / distribution", "title": "AWS releases TorchNeuron 2.33.0 for Trainium as a public beta", "whatChanged": "AWS released TorchNeuron 2.33.0 as a public beta on 20 October for Trainium2 and Trainium3. TorchNeuron registers Trainium as a native PyTorch device and supports eager execution, torch.compile, standard distributed APIs and custom kernels through NKI. Hugging Face recommends Transformers, Diffusers, Accelerate, PEFT and TRL directly on Trainium.", "whyCare": "You can run standard PyTorch training and inference code on Trainium, which gives teams a new deployment option alongside GPUs and hosted model APIs.", "before": "Teams ran PyTorch XLA with Optimum Neuron, which maintained Neuron-specific model code and trainers.", "after": "TorchNeuron exposes Trainium through PyTorch's native device backend, where standard models can use eager execution, torch.compile and distributed APIs on Trainium2 and Trainium3.", "learn": "Check supported models, operators and Ubuntu versions before moving workloads. Test the public beta, and confirm your environment supports Ubuntu 26.04, the target for current wheels.", "cloudPrimitives": [ "Compute", "Networking", "Storage" ], "sources": [ { "name": "PyTorch Blog - Introducing TorchNeuron: Native PyTorch backend for AWS Trainium", "url": "https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium/" }, { "name": "AWS Neuron - TorchNeuron 2.33.0 release notes", "url": "https://github.com/aws-neuron/torch-neuronx/releases" }, { "name": "Hugging Face Blog - TorchNeuron: running HF models on AWS Trainium is easier than ever", "url": "https://huggingface.co/blog/hf-torchneuron" }, { "name": "AWS Neuron documentation - What is TorchNeuron?", "url": "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/torch/about-pytorch-neuron.html" } ], "architectureImplication": "TorchNeuron lets standard PyTorch APIs run on Trainium. Compare framework coverage, compilation behavior, hardware availability and total serving cost before moving workloads.", "tradeoffs": [ "Compatibility", "Cost", "Complexity", "Reliability" ], "vendors": [ "AWS" ], "decisions": [ "deployment" ] }, { "id": "2026-10-20-mistral-voxtral-3", "date": "2026-10-20", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "real-time voice agent stack", "impact": "Packaging / distribution", "title": "Mistral launches Voxtral 3 for real-time voice agents", "whatChanged": "On 20 October, Mistral launched Voxtral 3, a voice-agent SDK that streams speech recognition, model reasoning, tool calls, and speech synthesis in one pipeline. Teams can use Mistral's EU-hosted API or deploy the models on their own infrastructure.", "whyCare": "You can build spoken agents without wiring separate transcription, reasoning, tool-use, and speech services. Mistral offers an EU-hosted API and self-deployment, giving you choices for audio processing and control.", "before": "Teams connected speech recognition, reasoning, tools, and speech synthesis as separate stages.", "after": "Voxtral 3 combines listening, reasoning, tool calling, and speaking in a real-time pipeline. Teams can use hosted or self-deployed models.", "learn": "Compare latency, language coverage, transcription quality, deployment needs and cost with OpenAI's real-time audio offerings and separate speech stacks. Test audio retention and tool-action controls before using Voxtral in customer workflows.", "cloudPrimitives": [ "Compute", "Networking", "Storage", "Security" ], "sources": [ { "name": "Mistral - Voxtral 3", "url": "https://mistral.ai/news/voxtral-3/" }, { "name": "Mistral documentation - Audio overview", "url": "https://docs.mistral.ai/studio/audio/overview" } ], "architectureImplication": "Design the workflow around streaming audio, turn detection, tool permissions, and retention policy. Compare the integrated pipeline with separately replaceable speech and language services before you commit to it.", "tradeoffs": [ "Accuracy", "Latency", "Cost", "Control", "Portability" ], "vendors": [ "Other" ], "decisions": [ "agent-vs-workflow", "deployment", "build-vs-buy" ] }, { "id": "2026-10-20-google-meta-torchtpu", "date": "2026-10-20", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "native pytorch tpu backend", "impact": "Packaging / distribution", "title": "Google and Meta introduce TorchTPU as a native PyTorch backend", "whatChanged": "On 20 October, Google and Meta introduced TorchTPU, an open-source PyTorch backend for Google TPUs. It runs eager and torch.compile workloads and supports distributed training. The project reports that 5,000 popular Hugging Face models run out of the box.", "whyCare": "TorchTPU lets PyTorch teams evaluate TPU deployments in PyTorch. Since it remains in active development, check operator coverage and workload performance.", "before": "PyTorch teams ran TPU workloads through PyTorch/XLA or other accelerator-specific integrations.", "after": "TorchTPU exposes TPUs as a PyTorch device and integrates with torch.compile, vLLM, SGLang, TorchTitan, and Ray.", "learn": "Test your models, distributed patterns, supported operations, and compile behavior on TorchTPU. Compare its performance and total cost with PyTorch/XLA, GPUs, and AWS TorchNeuron.", "cloudPrimitives": [ "Compute", "Networking", "Storage" ], "sources": [ { "name": "PyTorch Blog - Introducing TorchTPU", "url": "https://pytorch.org/blog/introducing-torchtpu/" }, { "name": "Hugging Face - TorchTPU: TPUs, Natively", "url": "https://huggingface.co/blog/hf-torchtpu" }, { "name": "Google and Meta - TorchTPU project", "url": "https://github.com/google-pytorch/torch_tpu" } ], "architectureImplication": "Evaluate TPU as a first-class PyTorch deployment target. Keep backend and hardware-specific tuning behind an interface. TorchTPU is in active development, so a simple device change does not guarantee peak performance.", "tradeoffs": [ "Compatibility", "Cost", "Control", "Complexity", "Performance" ], "vendors": [ "Google", "Meta" ], "decisions": [ "deployment", "cost-and-routing" ] }, { "id": "2026-10-22-aws-sagemaker-training-sessions", "date": "2026-10-22", "layer": "Model", "kind": "signal", "changeType": "serverless custom training", "impact": "Incremental", "title": "AWS makes SageMaker AI training sessions generally available", "whatChanged": "AWS made SageMaker AI training sessions generally available on 22 October. Stateful, serverless GPU sessions expose sampling, gradient and optimizer calls for custom supervised fine-tuning or reinforcement-learning loops. AWS manages GPU allocation, distributed execution, fault recovery and checkpoints for four supported models.", "whyCare": "You can run custom loss functions, reward signals and multi-turn reinforcement-learning loops without operating GPU clusters. Check supported models and Regions before moving an experiment.", "before": "Teams used predefined SageMaker customization recipes or managed their own GPU clusters for custom training loops.", "after": "Teams can compose stateful SageMaker AI session calls for sampling, gradient updates and optimization. AWS allocates GPUs and manages distributed execution, fault recovery and checkpoints.", "learn": "Check supported models and Regions. Compare per-token charges with GPU job costs. Test recovery, checkpointing and model quality on your reward loop.", "cloudPrimitives": [ "Compute", "Storage", "Security" ], "sources": [ { "name": "AWS Machine Learning Blog - Announcing new ways to customize models with SageMaker AI, without managing GPUs", "url": "https://aws.amazon.com/blogs/machine-learning/announcing-new-ways-to-customize-models-with-sagemaker-ai-without-managing-gpus/" }, { "name": "AWS SageMaker AI - Training session APIs", "url": "https://docs.aws.amazon.com/sagemaker/latest/dg/customizing-models-training-sessions.html" } ], "architectureImplication": "Keep reward logic, evaluation and loop control in your code. Call SageMaker AI for GPU-backed sampling and weight updates. Validate data movement, session quotas and cost against self-managed training jobs.", "tradeoffs": [ "Cost", "Control", "Portability", "Complexity" ], "vendors": [ "AWS" ], "decisions": [ "managed-vs-custom-runtime", "deployment", "cost-and-routing" ] }, { "id": "2026-10-23-github-copilot-client-usage-metrics", "date": "2026-10-23", "layer": "Production & Business", "kind": "signal", "changeType": "client usage reporting", "impact": "Incremental", "title": "GitHub adds client-level Copilot usage metrics", "whatChanged": "On 23 October, GitHub added `totals_by_client` to Copilot usage reports for enterprise, organization, and user scopes. It reports model requests, sessions, prompt and output tokens, and client breakdowns in one-day and 28-day reports. The Copilot usage metrics policy must be enabled, and client IDs are best-effort.", "whyCare": "You can compare Copilot model use across IDEs and external tools, then use request and token counts to guide support and governance. Client labels reflect integration IDs, so verify them against your own identity and endpoint telemetry.", "before": "Reports showed usage by IDE but missed requests from tools without editor telemetry.", "after": "Reports now group model requests, sessions, prompt tokens, output tokens, and client totals across one-day and 28-day views.", "learn": "Enable the Copilot usage metrics policy. Compare client labels with your own telemetry before using them for attribution or governance.", "cloudPrimitives": [ "Storage", "Security" ], "sources": [ { "name": "GitHub Changelog - Usage metrics API adds usage from tools like OpenCode", "url": "https://github.blog/changelog/2026-10-23-usage-metrics-api-adds-usage-from-tools-like-opencode/" } ], "architectureImplication": "Add `totals_by_client` to Copilot usage reviews. Treat the client value as a best-effort grouping and verify it against identity and endpoint telemetry before using it for policy decisions.", "tradeoffs": [ "Attribution", "Visibility", "Governance" ], "vendors": [ "GitHub" ], "decisions": [ "observability-and-evals", "cost-and-routing" ] }, { "id": "2026-10-26-github-copilot-assisted-permissions-ga", "date": "2026-10-26", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "model-assisted permission checks", "impact": "Incremental", "title": "GitHub makes assisted permissions generally available across Copilot", "whatChanged": "GitHub made assisted permissions generally available in Copilot CLI, the Copilot app and VS Code. A model evaluates command, file-path and URL requests. It lets requests it judges clearly safe proceed, while users must approve others. Enterprise admins can set defaults, block Allow All or disable assisted mode.", "whyCare": "You can reduce routine approval prompts while keeping human review for requests the model does not classify as clearly safe. Admins can keep assisted mode available while blocking Allow All.", "before": "Assisted permissions were not generally available. Developers could use manual approval or Allow All.", "after": "Copilot now evaluates requested commands, file access and URL opens. It runs requests it judges clearly safe and asks for approval on others.", "learn": "Pilot assisted mode on representative developer tasks. Review which requests proceed without approval. Set an enterprise default and decide whether policy allows Allow All.", "cloudPrimitives": [ "Security", "Policy", "GitHub Copilot" ], "sources": [ { "name": "GitHub Changelog - Copilot assisted permissions generally available", "url": "https://github.blog/changelog/2026-10-26-github-copilot-assisted-permissions-generally-available" } ], "architectureImplication": "Pilot assisted permissions on representative developer tasks and validate which operations proceed without review. Keep policy controls and sandboxing alongside the model's decision.", "tradeoffs": [ "Approval latency", "Security", "Control" ], "vendors": [ "GitHub" ], "decisions": [ "identity-and-permissions", "human-approval", "security-and-guardrails" ] }, { "id": "2026-10-27-github-copilot-sdk-server-auth", "date": "2026-10-27", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "workload identity", "impact": "Incremental", "title": "GitHub lets organization apps authenticate Copilot SDK agents", "whatChanged": "GitHub added public-preview server-to-server authentication for the Copilot SDK on 27 October. Organization-installed GitHub Apps can use short-lived installation tokens without a signed-in user. GitHub attributes usage to the organization and bills it there. The organization applies its model policies, budgets and spending controls. Owners can audit app usage and revoke access.", "whyCare": "You can run Copilot SDK services without a human account token. Your organization can apply its model policies and spend controls, while app-level usage reports show attribution and owners can revoke access.", "before": "Services and agents relied on a signed-in user's credentials, tying authentication and usage to people. GitHub Actions already offered a built-in GITHUB_TOKEN route.", "after": "An organization can grant a GitHub App the read-only Copilot Requests permission, then pass its one-hour installation token to the SDK. GitHub attributes usage to the app's organization.", "learn": "Test app-owned identity in a non-production service. Scope the GitHub App to the needed organization, rotate one-hour installation tokens, and verify usage reporting and revocation.", "cloudPrimitives": [ "Compute", "Security", "Storage" ], "sources": [ { "name": "GitHub Changelog - GitHub Copilot supports server-to-server authentication", "url": "https://github.blog/changelog/2026-10-27-github-copilot-supports-server-to-server-authentication" }, { "name": "GitHub Docs - Server-to-server authentication", "url": "https://docs.github.com/copilot/how-tos/copilot-sdk/auth/server-to-server-tokens" } ], "architectureImplication": "Give persistent SDK services a GitHub App installation identity. Scope installations so organization policies, budgets and audit logs govern usage.", "tradeoffs": [ "Security", "Control", "Cost" ], "vendors": [ "GitHub" ], "decisions": [ "identity-and-permissions", "cost-and-routing" ] }, { "id": "2026-10-28-aws-lambda-durable-execution-telemetry", "date": "2026-10-28", "layer": "Production & Business", "kind": "signal", "changeType": "execution telemetry", "impact": "Packaging / distribution", "title": "AWS Lambda exports durable-run telemetry and adds an MCP query server", "whatChanged": "AWS released the open-source Durable Execution Telemetry Exporter and a companion MCP server on 28 October. The exporter records each durable run's status, steps, waits, retries, timing, input and output, then sends a JSON record to one or more of 13 destinations. The MCP server queries TypeScript, Python and Java records read-only.", "whyCare": "You can query a durable workflow across Lambda invocations without joining invocation logs and traces by hand. An MCP client can inspect records without writing queries for each data store.", "before": "Lambda stored logs per invocation, splitting a durable run across streams. Teams had to aggregate records to analyze failures across runs.", "after": "The exporter emits one structured record per execution and routes it to selected stores. An MCP server lets agents query supported records read-only.", "learn": "Choose a destination and retention plan for run records. Redact inputs and outputs before export. Give the MCP server read-only AWS credentials.", "cloudPrimitives": [ "Compute", "Storage", "Security", "Networking" ], "sources": [ { "name": "AWS What's New - Lambda durable telemetry", "url": "https://aws.amazon.com/about-aws/whats-new/2026/10/lambda-durable-telemetry/" }, { "name": "AWS Compute Blog - Observing AWS Lambda durable functions with the Durable Execution Telemetry Exporter", "url": "https://aws.amazon.com/blogs/compute/observing-aws-lambda-durable-functions-with-the-durable-execution-telemetry-exporter/" }, { "name": "AWS Documentation - Durable Execution Telemetry MCP server", "url": "https://docs.aws.amazon.com/durable-execution/sdk-reference/observability/telemetry/mcp/" } ], "architectureImplication": "Treat run telemetry as a separate data product. Route records to stores that match your retention and query needs. Scope MCP access to read-only credentials.", "tradeoffs": [ "Security", "Cost", "Complexity", "Control" ], "vendors": [ "AWS", "Open standards" ], "decisions": [ "observability-and-evals", "integration-standards", "security-and-guardrails" ] }, { "id": "2026-10-29-aws-sagemaker-soci-images", "date": "2026-10-29", "layer": "Agent & Infrastructure", "kind": "signal", "changeType": "lazy-loaded container images", "impact": "Incremental", "title": "AWS starts SageMaker jobs from SOCI-indexed images in seconds", "whatChanged": "Amazon SageMaker AI added SOCI-indexed images for training and processing jobs on 29 October. SageMaker streams required files first, then fetches the rest while code runs. AWS says this cuts GPU idle time during image pulls by up to 90%. Jobs without an index still download the full image.", "whyCare": "Large container pulls can leave paid accelerators idle before work starts. You can use pre-indexed AWS Deep Learning Containers or index your ECR image with open-source tools. Job definitions stay the same.", "before": "SageMaker downloaded the complete image before running a job. Large deep-learning images could add minutes to each run.", "after": "SageMaker checks for a SOCI index, starts the container with required files, then streams the rest. Unindexed images still require full downloads.", "learn": "Benchmark large images on actual jobs. Index images in ECR with supported tools or use AWS's pre-indexed -soci tags. Test workloads that read many image files before training.", "cloudPrimitives": [ "Compute", "Storage", "Networking" ], "sources": [ { "name": "AWS What's New - SageMaker AI supports SOCI-indexed images", "url": "https://aws.amazon.com/about-aws/whats-new/2026/10/sagemaker-ai-soci-images-training/" }, { "name": "AWS Documentation - Speed up training job startup with SOCI-indexed images", "url": "https://docs.aws.amazon.com/sagemaker/latest/dg/train-soci-indexed-images.html" }, { "name": "AWS Machine Learning Blog - Announcing SOCI-indexed images for faster SageMaker AI jobs", "url": "https://aws.amazon.com/blogs/machine-learning/announcing-soci-indexed-images-for-faster-sagemaker-ai-jobs/" } ], "architectureImplication": "Add indexes to image artifacts without changing job definitions. Compare SOCI with full pulls. Watch for jobs that read most image files before startup, when savings shrink.", "tradeoffs": [ "Cost", "Latency", "Complexity" ], "vendors": [ "AWS" ], "decisions": [ "deployment", "cost-and-routing" ] } ], "overallDirection": { "title": "From assistants to agents and automated knowledge work", "summary": "The broad direction is toward systems that reason longer, use tools, act across software, and take on larger portions of real work. Model capability still matters, but the architectural bottleneck is moving toward reliable autonomy and integration.", "destination": "Increasingly general AI systems that can perform broad cognitive work, research, and operations under human direction. This is a direction of travel, not a guaranteed end state.", "constraint": "Alignment, monitoring, permissions, security, reliability, and preserving meaningful human control as systems become more autonomous.", "sources": [ { "name": "OpenAI on AGI in steps", "url": "https://openai.com/safety/how-we-think-about-safety-alignment/" }, { "name": "Automated research intern milestone", "url": "https://openai.com/index/research-acceleration-view-inside-openai/" }, { "name": "Anthropic frontier safety roadmap", "url": "https://www.anthropic.com/responsible-scaling-policy/roadmap" } ] } }