This Past Week in AI

Edition #3·August 22, 2026·Week of Aug 16–22

This edition resumes Saturday publication after a pause since 14 March 2026. The week of Aug 16–22 was not about a single model launch. It was about agents becoming infrastructure: durable surfaces in developer tools, a dedicated host process in VS Code, production computer use from Anthropic, and a clearer split between what models can do and what teams can actually contain.

GitHub Moves Copilot Agents Onto Durable Canvases

GitHub used the week to argue that chat is a weak place to run real agent work. In the Copilot app, canvases give developers and agents a persistent shared surface where workflow state, drafts, and approval gates stay visible instead of disappearing into a transcript. GitHub Developer Advocate Ayan Gupta published two examples — Java Modernization Studio and Site Studio — in the open-source awesome-copilot collection, and pointed teams toward starting small with /create-canvas on a recurring workflow. The useful signal for software teams is architectural, not cosmetic: once agents outpace human review, the product problem is no longer prompting. It is making progress inspectable, steerable, and approvable. Source

VS Code Treats the Agent Runtime as Its Own Process

Visual Studio Code 1.134, released August 19, continues that same shift inside the editor. Agent sessions now run in a dedicated Agent Host process based on the Agent Host Protocol, so the same session can span multiple windows and outlive a closed editor. The Copilot harness in that host is powered by the Copilot SDK, aligning behavior with Copilot CLI and the standalone Copilot app. The rest of the release is about navigating the work those sessions produce: side-by-side chats, a prompt timeline with file-change counts, and find-in-chat across the full transcript. For teams already running agents in the editor, this is Microsoft treating the agent runtime as infrastructure rather than an extension-host feature. Source

Anthropic Ships Computer Use to Production, With a Browser Tool

Anthropic moved computer use, the Skills API, and the Files API to general availability on the Claude Platform, and added a browser use tool for agents that work in web applications. Computer use can now take several actions per turn instead of one per model call, and Anthropic says it is eligible for HIPAA-regulated workloads under its BAA. The browser tool is the more important change for application teams: alongside a screenshot, Claude reads page structure and targets a specific field or button rather than guessing pixel coordinates. Skills let teams upload and version their own instructions and scripts; the Files API stores the documents agents read and write, with automatic expiration, higher rate limits, and 1 TB of storage per organization. Skills and Files are also available through Microsoft Foundry, with the updated computer and browser tools coming to Google Cloud's Vertex AI. The product implication is blunt: agents are being packaged to operate software that was never built for automation. Source

Google Publishes a Zero-Trust Pattern for Production Agents

Google's Agent Development Kit post made the security argument that should follow every computer-use launch. Once an agent can issue refunds, modify databases, or execute generated code, it is mutating production state — and system prompts are not a security boundary. Google open-sourced a customer-support agent demo and described three hard controls outside the model: hardware-backed cryptographic signatures on every write, gVisor sandboxing with no network egress for generated code, and a deterministic semantic gateway in front of model calls and tool output. The reference implementation maps onto Cloud KMS, gVisor, and VPC Service Controls in production. For teams doing practical AI integration with LLMs and agents, this is the production checklist hiding inside a vendor blog: identity, isolation, and deterministic policy have to live in infrastructure, not in the prompt. Source

Frontier Labs Still Won't Publish How They'd Contain a Rogue Model

The week's sharpest governance story came from Guidelight AI Standards, via TechCrunch. Few frontier labs have published a containment plan that says what permissions get revoked, who a misbehaving model may keep serving, and when it comes fully offline. Guidelight graded public materials from Anthropic, Google, OpenAI, Meta, and xAI; OpenAI scored highest, Anthropic and Meta lowest. Google and OpenAI both said the assessment does not capture their full internal practices. An OpenAI spokesperson said the company has a process for restricting permissions, pausing workloads, limiting deployment, or taking a model offline, and has used it. The gap still matters for anyone deploying agents inside real systems: California's SB 53 already requires large frontier developers to explain how they handle critical safety incidents, New York's RAISE Act takes effect in January, and Congress has a bipartisan kill-switch bill in play. Disclosure is not the same as a plan, but the absence of a public plan is now a procurement question. Source

From Our Perspective

The throughline this week is that agent capability is moving into the tools where software already gets built, while containment is still mostly a prompt and a hope. Copilot canvases and VS Code's Agent Host make long-running agent work visible. Anthropic is selling computer use as production infrastructure. Google is telling teams that identity, sandboxes, and deterministic gateways — not system prompts — are the actual control plane. That is the right problem set for software teams: not whether to use models, but how to put them in production without handing them the keys.