AGENTS.md Support, Server-Side Auto Mode, and Skill Sync

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of Monday, September twenty-first. Claude Code now reads AGENTS.md when there's no CLAUDE.md. Auto mode's classifier moved server-side. And the New York Times reports Anthropic is pursuing a public listing.

Claude Code now reads AGENTS.md when there's no CLAUDE.md. Auto mode's classifier moved server-side. And the New York Times reports Anthropic is pursuing a public listing.

Let's start with Claude Code, where eight releases this week landed with a clear theme -- configuration, and getting out of your way. The headline: if your repository already carries an AGENTS.md file for other coding agents, Claude Code will now read it in projects that have no CLAUDE.md. You don't have to maintain two instruction files anymore, and you can switch which one gets used under project instructions in the config screen. One caveat -- this isn't available yet on Bedrock, Vertex, or Foundry.

That same idea of not maintaining two of everything shows up in skills, too. Skills and plugins you've enabled on your claude dot A I account now appear automatically in terminal sessions signed in with that account. If you'd rather keep them separate, there are settings to turn each sync off individually. Staying on the cloud platforms for a moment -- if you use auto mode on the Claude API, Enterprise, Bedrock, Vertex, Foundry, or through a gateway, the safety classifier now runs on the server by default, and you're no longer charged for its token overhead.

If a billed fallback does happen, you'll get a warning, and there's a new auto mode server row in the status view so you can see which path your session is on. There's an opt-out if you need it. On the security side, plugins installed from an npm source are now fetched with install scripts disabled and integrity-verified, so a package's install scripts no longer execute on your machine. And rather than blanket-approving everything, you can now approve the exact command a previous run displayed to you.

Two workflow changes round it out. There's a new keyboard shortcut that interrupts what Claude is doing and pushes all your queued messages through at once -- and sent and queued messages stay gray until the model actually receives them, so you can tell what's landed. In Visual Studio Code, you can now accept or reject each change individually in the proposed-change diff tab instead of taking a whole edit at once. Background shells and other running tasks show up in the agent map with a stop control, and responses get a copy button.

One removal worth flagging: the TaskOutput tool is gone. Claude now reads a background task's output file with the standard read tool instead, and the old output-length settings no longer do anything.

Under the hood, this was a cleanup week, and most of it centers on the resume path. Sessions no longer crash at launch or on resume when your Claude config file holds a malformed theme value, a malformed authentication notice flag, or a malformed custom API key response. Same for saved transcripts containing a malformed message block. And that endless unexpected tool use I D retry loop -- the one that would spin forever -- now either self-heals or ends with a clear error.

On the performance side, large diffs and long transcripts render faster with fewer slow frames. Per-turn slowdowns caused by language server diagnostics and hook progress updates are gone. And headless sessions -- SDK and print mode -- no longer wait on the per-directory CLAUDE.md lookup before the first turn, so they start noticeably quicker.

Now to the broader Anthropic news, and it starts with a security story. The Wall Street Journal reports that attackers used Claude to assist an intrusion at OpenAI. Worth keeping in mind alongside this week's classifier changes -- expect continued tightening of the safety systems that sit in front of the models you call from Claude Code. That tension between capability and caution runs through the next story as well.

The New York Times reports Anthropic is moving ahead with a public listing, now slated for November, even as its leadership publicly argues for slowing frontier development. How that resolves shapes how quickly new model capabilities reach your terminal. Meanwhile, Reuters reports that Anthropic says Claude now drives roughly a quarter of the engineering work on its next models. That's a useful datapoint on how far agentic coding has gone inside the company that builds your tooling -- and it connects directly to something on the Claude blog this week.

Anthropic's continuous integration job volume grew twenty-five times in six months as agents wrote more code, and three separate patches to its test selection service failed before a durable fix held. If your own pipelines are buckling under agent-generated pull requests, the failure modes will look familiar. We'll link that one, plus a post on the redesigned projects experience, in the show notes. And finally, Axios reports Anthropic launched document tools and folded Cowork into the main Claude interface -- the same consolidation behind this week's Cowork and cloud session changes.

That's it for the brief. I'm Shannon, and we'll see you next week.