Microsoft Research and GitHub Next’s PDD Kit turns plain-language intent into a Verus spec, proves agent-written Rust against it, and labels what’s only tested. Here’s how the Pet Store login demo works and where the proof stops.
OpenRouter rebuilt its Auto Router and launched Auto Code, a router tuned for coding harnesses that picks a model per turn, tries not to break your cache, and switches when tests keep failing. Here’s what it claims and what I’d check.
From Mexico: Kiro just made bots GA in Kiro Crew — named, long-running agents with their own memory, scoped permissions and schedules that run parallel sessions for you. Here’s what it changes, and what I’d test before trusting it.
From Mexico: Google says Antigravity with Gemini 4 Argon rewrote a 174k-line Fortran coastal ocean model core into safe Rust in 59 minutes, caught old bugs, and open-sourced the skill. Here’s what the run claimed — and what I’d verify.
From Mexico: GitHub Copilot now runs inside JetBrains Air over the Agent Client Protocol, locally or in sandboxed cloud tasks on your Copilot plan. Here’s what works, what Air can’t see or limit, and why the protocol matters more.
From Mexico: OpenAI says Codex pushed its monorepo past what its Git servers could handle, so it built Pando, a Git-compatible backend that lets agents push to many branches in parallel. Here is how it works and what I take from it.
From Mexico: DeepSeek open-sourced dsh-superagent, a research prototype coding agent that listens to GitHub, takes requests only from people you allow, works each in its own session and folder, and answers with a pull request.
From Mexico: Supabase shipped Explorer and Notebooks, one editor for database and log queries plus saved, rerunnable investigations that live in your repo. Its Assistant can write the cells, but only with your approval.
From Mexico: Hud launched agentic rollouts, which check an agent’s PR against how the code really runs in production, watch the canary at the function level and roll back fast. Here’s how it works and what I’d verify first.
From Mexico: Adversa AI launched a security platform that watches what coding agents like Claude Code and Codex actually do on developer machines and blocks the risky step in a chain. Here’s what it covers and what I’d check.