Google's DeepMind reshuffle spins off a new AI lab overnight, OpenAI cuts prices while solving decade-old math problems, Claude Code closes four permission bugs, and 80 Acres Farms becomes indoor agriculture's biggest casualty yet.
A Claude Code default let an AI agent act without asking after 60 seconds of silence. The same pattern runs today, unexamined, in how AI assistants handle memory.
A live npm supply chain attack meets GitHub's own hardening push, MCP drops its session handshake, Anthropic stands alone on open weights, and agriculture tech sees four distinct developments from funding to hatchery automation.
GitHub pairs a Claude Opus 5 rollout with new agent oversight controls in Issues, two arXiv papers find agent skill libraries backfire as often as they help, Kimi K3 tops Hacker News, and agriculture capital keeps favoring biology over hardware.
An open-weights model tops Hacker News, GitHub folds AI security detection into code review, an adoption study finds agentic coding still runs through one reviewer, and farm autonomy proves out on retrofit economics rather than novelty.
TypeScript 7's native compiler, causal audits for coding agents, Copilot's expanding control plane, and an investor audit of regenerative agriculture claims.
A signal catalog for the pipeline itself: which stage emits each signal, who consumes it, and what to alert on. Standardized where it can be, honest where it is not.
Co-authored-by is the wrong trailer for AI: it is human credit, and tools get it wrong. Use a dedicated Assisted-by trailer, carried into signed provenance.
How an attested Claude Code plugin marketplace verifies, signs, and SHA-pins releases, and why none of it replaces deciding your own upgrades yourself.
Coding-agent evaluation, Copilot governance, Claude model policy, secret exposure, API access limits and soil sensing finance defined the developer week.