Summary

Today's six signals trace a maturing agent stack pulling in three directions at once. Anthropic graduated computer use, the Files API, and Agent Skills out of beta and launched a dedicated browser-use tool, signaling its agent tooling is production-grade. At the same time, agents are spilling into new surfaces: GitHub put Copilot into Slack in public preview, and KTX shipped a Skills Kit that lets agents in Claude, ChatGPT, Codex, and Cursor place live trades. The countervailing force is control and proof: Nuggets launched an in-path Authority Control Plane to authorize and revoke agent actions, while Microsoft's ThinkingBox and Google's EnvHarness bring outcome-based benchmarking and programmable training environments that expose how unreliable agents still are under repetition.

Key themes

  • Agent tooling matures to production: Anthropic took computer use, Files, and Agent Skills out of beta and added a purpose-built browser-use tool plus domain allow/block controls, treating its agent stack as production-grade rather than experimental.
  • Agents move into the surfaces where work happens: GitHub Copilot entered Slack in public preview and KTX's Skills Kit wired trading intelligence and live order execution into Claude, ChatGPT, Codex, and Cursor, pushing agents from IDEs and standalone apps into chat and assistant surfaces.
  • Governance and authorization become first-class infrastructure: Nuggets' Authority Control Plane sits in the execution path to evaluate identity, delegated authority, policy, and intent per action, with real-time revocation and a fail-closed kill switch, as agents begin initiating transactions.
  • Evaluation and training get harder and more honest: Microsoft's ThinkingBox grades agents on real database state changes across repeated runs (65% pass@1 collapsing to 25% over 20 runs), and Google's EnvHarness makes static training environments programmable to target agent weaknesses.

Notable items

  • Anthropic takes computer use, Files, and Agent Skills out of beta on the Claude Developer Platform and launches a browser-use tool (browser_toolset_20260801) for driving a hosted browser, plus allow/block domain lists for managed-agent web search and fetch. (High impact, primary source)
  • Microsoft open-sources ThinkingBox and ThinkingBox-Bench, an MCP-compatible sandbox that scores agents on backend state changes across 507 policy-conditioned workflows; the best of 12 models hit 65% pass@1 but only 25% pass^20, reframing reliability under repetition.
  • Nuggets launches the Authority Control Plane, a standalone, cross-system in-path layer to define, delegate, enforce, and prove agent authorization with real-time revocation and an estate-wide fail-closed kill switch, sharpened by EU AI Act high-risk obligations.
  • GitHub launches Copilot in Slack in public preview, bringing Copilot CLI and app agent capabilities (planning, issue investigation, bug triage, PR creation, async hand-off) into chat under existing GitHub permissions.
  • KTX launches its Skills Kit in Seoul, distributing market data, AI-generated intelligence, portfolio tools, and live order execution as agent skills into Claude, ChatGPT, Codex, and Cursor.
  • Google Research open-sources EnvHarness (arXiv 2608.19880), a programmable layer that reshapes static agent training environments without rewriting their logic, reporting a 9-point gain on out-of-distribution tasks.

Source coverage

Source rows used: 6