A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
Worktree-isolated sessions were never sealed off from your main checkout
Version 2.1.222 fixes worktree-isolated sessions and their subagents being able to run destructive git commands against the main checkout; isolation now applies to file edits and Bash in every session type. The same release stops PreToolUse auto-allow hooks from bypassing tool restrictions in background agent tasks such as summaries, compaction and renames. Messages sent between agent sessions now pass the permission classifier before dispatch, and repo-local settings can no longer switch Remote Control on.
An open framework now trains coding agents inside the harness they will actually run in
Microsoft Research released Orchard, an open-source framework built on a Kubernetes-native environment service that provides sandbox lifecycle primitives reusable across task domains, agent harnesses and training stages. Its software-engineering recipe, Orchard-SWE, distils 107K trajectories and reaches 69.7% on SWE-bench Verified — 73.0% with value-model reranking — from roughly three billion active parameters. Microsoft is publishing the code, training data and evaluation methods, and the same service backs software-engineering, browser and personal-assistant agents without modification.
Your Copilot and Claude Agent SDK agents now answer to one identity and safety policy
Microsoft's Agent Harness and Foundry Hosted Agents reached general availability, four months after the framework itself hit 1.0 in April. The harness wraps a model with function invocation, per-call history persistence, context compaction and planning modes, file memory and built-in OpenTelemetry observability; shell tooling and background sub-agents are available but carry warnings. Two coding-agent connectors shipped with it, one for the GitHub Copilot SDK and one for the Claude Agent SDK, so external agents can be orchestrated under a single identity, content-safety and observability policy.
The llm CLI prints reasoning traces to stderr and calls provider-side tools
LLM 0.32 shows reasoning models' traces on standard error, so thinking is visible without polluting piped output, and -R/--hide-reasoning turns that off. Server-side provider tools are now callable directly, and the companion llm-anthropic 0.26 adds WebSearch, WebFetch, CodeExecution and AnthropicMCP for the Claude 5 models. Logging moved to a content-addressable message store, cutting duplicated JSON while keeping llm logs readable.
