Claude Code’s AGENTS.md fallback sat behind a remote feature flag that failed when telemetry or nonessential traffic was disabled, so the local file was skipped with no warning. An Anthropic engineer said the gate was a rollout kill-switch and is fixed in v2.1.281.
GitHub’s Copilot app now supports per project local sandboxing (public preview): filesystem allow/deny lists, outbound/local network rules, and optional Git/GitHub CLI credentials. It’s off by default; turn it on in project settings or with /sandbox on , and sessions fail closed if the OS can’t enforce the policy…
Anthropic’s system card shows an early Opus 5.5 snapshot acted on instructions planted in user pasted text in 52% of coding tests (26% for invisible Unicode). The released model drops to 2% at default effort and 0% with product mitigations; Claude Code now strips invisible characters and marks pasted text…
Meta Superintelligence Labs product head Nat Friedman said Muse was “built from scratch” but “definitely heavily inspired” by OpenClaw after users found matching workspace filenames and nearly identical SOUL.md content. He said the team thought OpenClaw creator Peter Steinberger “got those things exactly right” and wanted an OpenClaw like agent that could scale to billions…
Unreal Labs open sourced Unreal Agent, a Go harness that runs tool calls asynchronously so agents skip poll/wait token tax. On Terminal Bench 4.0 with GPT 6 Astra xhigh it matched Codex at 57.9% for $1,428 vs Codex’s $2,350, and claims up to 40% cost savings vs Codex and 20% vs Pi on production workloads…
Simon Willison’s CLI tool llm 0.36 adds OpenAI model IDs gpt 6 sol and gpt 6 luna , plus a supports conversation = False flag for single turn models (first used by llm typesafe). Reasoning traces in llm logs Markdown now wrap in details tags…
Spotted something we missed? Start a thread.
