Claude Code 2.1.277 now reads AGENTS.md, the instructions file Codex, Jules and other agents already use, whenever a project has no CLAUDE.md.
Claude Code now reads AGENTS.md, the open instructions file most other coding agents already read, after more than a year of requests. Microsoft used AI agents to rewrite the GitHub Copilot runtime into more than 800,000 lines of Rust, measured 15.9x faster on one benchmark, and listed what the agents got wrong along the way. Google engineers open-sourced AX, software for running fleets of AI agents, each in its own sandbox. Z.ai apologized for its coding app ZCode up users' git history without asking, and open-sourced the app. And TypeSafe opened Jev to everyone while a rival builder says he published the same idea a year earlier.
Claude Code and the tools that build it #
Claude Code adopts AGENTS.md, the shared format rivals already read
Claude Code now reads AGENTS.md whenever a project has no CLAUDE.md, its own instructions file. AGENTS.md is the shared format Codex, Jules, Aider, Zed, Warp, VS Code and Devin already read. This is the first version to read it: 2.1.277, shipped Thursday, announced the same day by Anthropic engineer Thariq. The HN reaction was less excitement than relief: one top comment counted "a year and a half of everyone complaining," another said the change means deleting one-line CLAUDE.md files that existed only to point at AGENTS.md. If your team already writes AGENTS.md for Codex or Jules, Claude Code reads it too now, no duplicate file required.
Microsoft's agents rewrote Copilot's runtime in Rust, 15.9x faster on one test
Microsoft used its own Copilot agents to rewrite the GitHub Copilot runtime from TypeScript into more than 800,000 lines of Rust, shipped as 128 pull requests rather than one big switch. On one benchmark, Rust ran 15.9x faster and used a tenth of the memory, though Microsoft's Stephen Toub is explicit that's one workload, not universal. The honest part: "dozens" of regressions shipped along the way, mostly correctness bugs, all since fixed. His verdict: Rust caught bugs, but "if it compiles, it's correct" is a joke.
New agent infrastructure #
Google engineers open-source AX to run fleets of AI agents, each in its own sandbox
Google engineer Jaana Dogan announced AX, open-source software for running large numbers of AI agents at once; its GitHub page calls it Google's "open agentic orchestrator." Each job runs in its own sandbox, limited to the repos and websites it is allowed to reach, and idle jobs and resume in under a second. The site says one cluster can hold "billions" of them; nobody outside Google has tested that. No official Google channel has announced it, and the repo warns of "major breaking changes." One to watch, not to adopt.
Follow-ups: a security story and a speed claim #
Z.ai apologizes for ZCode's secret uploads, open-sources the app
Z.ai apologized for its coding app ZCode silently up users' git history, which a researcher exposed last week, and open-sourced the app. Two outside assessors, the government-backed China Academy of Information and Communications Technology (CAICT) and the security firm NSFOCUS, say the cloud storage that held the uploads is now empty. The researcher, ferstar, says the upload code is gone but whether older copies were deleted "remains unverifiable externally." A full security report is coming; it isn't out yet.
Jev drops its waitlist as a rival builder claims he had the idea first
TypeSafe dropped the waitlist on Jev, its "decision model" that returns typed answers instead of text. Its homepage now cites "193.6x Faster" for one workflow; founder Diogo Almeida claimed 20-200x faster than LLMs at launch. The same week, Nandakishor Mukkunnoth published Laya, an open decision-model system, citing papers he says predate Jev's approach by a year and noting Jev shipped "without technical papers, without open weights." We found no explanation of the gap from TypeSafe, nor any response to Mukkunnoth.
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
Also worth your time #
- Browserbase says Stagehand, its browser-automation kit for agents, runs scripts 2x faster on its cloud browsers than on "Playwright cloud equivalent browsers" — the vendor's own claim, from its README, with no independent benchmark yet.
- Victor Taelin says Bend 2, his proof-checked programming language, is no longer taking outside pull requests while he raises venture funding — his launch post has passed 1.6M views; he has not answered developer Liam Powell's measurement that the launch demo needed 442 lines of proof for 58 lines of rules.
The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.