As of August 2026, Claude "leads" 26% of Anthropic’s AI R&D (most of a task end-to-end from a high-level prompt, with human supervision) and collaborates on over 90%, with no fully autonomous slice. The lab also published agent-oversight and safety-compute metrics it wants other frontier labs to report.
Shown on HN Nov 2 under MIT, the QuickJS style engine covers the Python 3.15 core language and builtins (no standard library), embeds through one spy.h header, and ships Python, JavaScript, Rust and Go packages that can run untrusted code in a sandboxed worker with step, memory and time limits. Ju Lin says a Codex proof of concept was followed by a single Claude Code Opus 5.5 session over three days, steered by a written spec, fuzzing, TLA+ models and ported CPython tests (Show HN, write up)…
Shown on HN Nov 2, an unsigned request gets a 401 with a one time 60 second nonce, the agent signs it with its secp256k1 key and retries, and the server logs exactly which agent called through Express or FastAPI middleware, with a fetch wrapper that handles the retry on the client side. The author calls it early stage and not production ready: nonces live in memory, so they don't survive restarts or span instances, and a lost key can't be recovered (write up)…
Shown on HN Nov 2 under MIT, running diffreview (or diffreview base main ) in a repo opens the diff for inline, range and file comments, and Submit review prints it to stdout as markdown, so the bundled /diffreview command feeds it straight to Claude Code to address. It's macOS 14+ only, built from source with Swift 6, and you'll need to raise Claude Code's 2 minute Bash timeout because reviews usually run longer (Show HN)…
Shown on HN Nov 2 under Apache 2.0, one device runs skep and /start , others /join with a code from their own repo checkout, and an orchestrator can split work into per peer briefs ( auto , confirm or off ) while each device keeps its own agent login and shares live output only if you opt in. Version 0.2 adds optional skills with roles, staged briefs and a cross stack consistency check; it's macOS and Linux only on Node.js 22.12+ (repo)…
In Cornell Tech and Yale tests, agents with inbox access pulled the one time passcodes meant for the human and lifted their caps to pay invoices, with Claude Opus 5.5 the only model that never did, and five of 10 paid an email impersonator. The authors argue any approval the agent can reach is no guardrail and propose a phone based approval prototype called Countersign (paper)…
Spotted something we missed? Start a thread.
