Why it mattersTeams that maintain both a CLAUDE.md and an AGENTS.md so different coding agents each find their file can now stop, and projects that only shipped AGENTS.md work in Claude Code with no extra setup.
Anthropic released Claude Code v2.1.277 on 18 September at 18:06 UTC. In a project with no CLAUDE.md, Claude Code now reads AGENTS.md instead, and the file name can be changed under "Project instructions" in the /config menu. The changelog notes that the new behaviour is not yet available on Bedrock, Vertex or Foundry.
Why this file name matters
AGENTS.md started as the convention OpenAI's Codex uses to keep a project's coding-agent instructions in one place, and it spread to Cursor, Aider and a number of smaller CLIs. Claude Code has always looked for CLAUDE.md instead, which is why many open-source repositories carry both files with identical or near-identical contents.
Projects that only ship AGENTS.md now work with Claude Code with no extra setup. Projects that already ship CLAUDE.md are unchanged, because Claude Code still prefers CLAUDE.md when both exist. AGENTS.md is a fallback the tool falls through to when CLAUDE.md is absent.
Other visible changes in this release
The gateway path picked up two options for teams whose only outbound network path is a forward proxy. The new environment variable CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 makes every outbound request hand the proxy the hostname rather than resolving it locally, so a proxy that expects to see hostnames in the request line works as intended. Upstreams in a Claude apps gateway config also accept a headers: map, so a static header can be attached to every request that goes to a provider through a proxy in front of the gateway.
A background task's final update now shows a "waiting" line when it finishes while a panel like /tasks is open, so a completed task no longer sits invisible until the panel is closed.
The fixes that could bite a running team
Two of the fixes affect people who saw failures that looked opaque. claude -p and Agent SDK sessions that hit an internal error could hang with no result; they now report the error and exit with code 1, so a wrapper script can tell that something went wrong. Conversations that failed every request with "text content blocks must be non-empty" after an earlier assistant turn produced an empty text block alongside other content, including after --resume, now go through.
A separate fix restores logins that were being ended by an older Claude Code build running on the same machine, for example the CLI bundled inside a VS Code extension. On a machine with two versions installed, the older one no longer signs the newer one out.
Two rendering fixes are worth noting for anyone using long-running sessions: a rare failure that stopped the screen updating for the rest of the session after an internal rendering error, and a rare Windows case where a turn ended with "Out of memory" right after Claude replied, so the reply's tool calls never ran.
The full changelog lists roughly 40 fixes covering plugin installation, Remote Control sessions, the Edit and Write tools, and /mcp and /plugin manage panels.
Source
Source: Anthropic
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
