Claude Code now reads AGENTS.md when no CLAUDE.md is present

Claude Code adds an AGENTS.md fallback

Claude Code, Anthropic's terminal-based coding agent, now reads AGENTS.md as its source of project instructions when a repository has no CLAUDE.md. The change appears in the newest entry of the official Claude Code changelog and was highlighted on the Hacker News front page.

How the fallback works

According to the changelog, the behavior is a fallback rather than a replacement. Projects that already ship a CLAUDE.md keep working exactly as before; only when that file is absent does the agent look for AGENTS.md instead. Developers can review or change which file counts as the project's instruction source under the "Project instructions" setting in the /config menu.

There is one deployment caveat: the changelog notes that the feature is not yet available on the Bedrock, Vertex and Foundry variants of Claude Code, so teams running the tool through those enterprise channels will not see it yet.

Toward one instruction file per repository

AGENTS.md is an emerging convention for repository-level agent instructions: a plain Markdown file, typically at the repo root, that documents build commands, code style and constraints an AI assistant should respect. Its main draw is portability. Repositories supporting several assistants have often ended up maintaining one bespoke file per tool, such as CLAUDE.md for Claude Code or .cursorrules for Cursor, with the same guidance duplicated and prone to drifting apart.

By honoring AGENTS.md, Claude Code lets a repository keep a single canonical instruction file and skip maintaining a CLAUDE.md copy purely for one tool. Because the native file still wins when present, the change is backward compatible: nobody has to migrate, but new projects and multi-tool teams can standardize on AGENTS.md from the start.

What else landed in the release

The same changelog entry carries a batch of other changes. A new environment variable, CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1, targets Claude apps gateway setups whose only outbound path is a forward proxy: with it set, outbound requests hand the hostname to the proxy rather than resolving it locally. Gateway upstreams can also now declare a static headers map for a proxy running in front of a provider.

On the fix side, the release is extensive. Headless runs launched with claude -p, along with Agent SDK sessions, no longer hang without a result after an internal error; they now report the error and exit with code 1. A bug that logged users out when an older bundled CLI, for example one shipped inside an IDE extension, ran alongside the current build was fixed. Numerous crashes were also addressed, including ones triggered by malformed values in ~/.claude. and prompts containing terminal color codes, alongside a long list of plugin, sandboxing and rendering fixes.

Why it matters

Instruction-file fragmentation has been a quiet tax on maintainers: every tool wanted its own file, and keeping them in sync was manual work. A fallback read of AGENTS.md from one of the field's most prominent coding agents pushes the ecosystem toward a single shared convention, where repository instructions are written once and honored regardless of which assistant a developer opens.

The design choice matters as much as the feature. Anthropic picked compatibility over migration, with CLAUDE.md still taking precedence, which lowers the cost of adoption to zero while making AGENTS.md a safe default for new repositories. The near-term gap is enterprise deployments on Bedrock, Vertex and Foundry, which will have to wait for the feature to reach their channels.