Claude Code now reads AGENTS.md if there is no Claude.md

Anthropic updated Claude Code to recognize AGENTS.md as a project-level instruction file, activating the fallback automatically when no Claude.md is present in the repository root. The change lands in the Claude Code changelog without a version pin or release date, but it is live for all users running a current Claude Code installation.

This article covers exactly what changed, how the fallback logic works, and what developers already using AGENTS.md for tools such as OpenAI Codex or other multi-agent frameworks should verify before relying on the new behavior in production.

What shipped in Claude Code now reads AGENTS.md if there

Claude Code previously treated Claude.md as its sole source of project-level instructions — the file that tells the agent about code style, testing commands, off-limits paths, and any other standing context the repo owner wants enforced. If Claude.md was absent, the agent started from a blank slate regardless of what other instruction files existed in the tree. With this change, Claude Code checks for AGENTS.md when it finds no Claude.md, and it loads that file as if it were the primary instruction source.

The new behavior is a conditional fallback, not a merge. When Claude.md is present, AGENTS.md is not read; Claude.md continues to be the authoritative file. The addition therefore carries no risk of conflict or double-instruction for teams that have already written Claude.md. Teams that have not yet written one — or that have been maintaining AGENTS.md for a different tool — will now see that file honored automatically inside Claude Code sessions.

What improved in Claude Code now reads AGENTS.md if there

The practical improvement is reduced duplication for teams maintaining multiple AI coding agents simultaneously. Before this change, a team using both Claude Code and a Codex-compatible agent had to maintain two separate instruction files even when those instructions were identical. After this change, a single AGENTS.md serves both, as long as no Claude.md exists.

No benchmark figures or latency numbers accompanied the changelog entry; the improvement is purely operational — one file maintained instead of two for qualifying projects.

What you gain from Claude Code now reads AGENTS.md if there

Developers who adopted AGENTS.md as a convention across their toolchain — common in monorepos that run several agents in parallel — gain immediate compatibility with Claude Code without any migration work. The instruction file they already wrote for Codex or another agent will now also govern Claude Code's behavior in those repos, enforcing the same testing commands, lint steps, and off-limits directories across every agent that touches the codebase.

For solo developers or smaller teams, the change lowers the barrier to structured agent instructions. AGENTS.md has become a de facto standard filename in the open-source community because multiple tools recognize it. Anthropic joining that convention means a developer can pick up Claude Code for an existing project and find it already context-aware, without needing to know that Anthropic's original filename was Claude.md.

How to get Claude Code now reads AGENTS.md if there

The feature requires no opt-in flag or configuration change; it activates automatically once Claude Code is updated to the current release. To update an existing installation, run either of the following:

npm install -g @anthropic-ai/claude-code

claude update

After updating, place AGENTS.md in the repository root and confirm there is no Claude.md present in the same directory. Start a session normally:

claude

Claude Code will load AGENTS.md at session startup. To verify the file was picked up, check the opening context summary in the terminal; Claude Code lists loaded instruction files at the start of each session. No model-switch or launch flag is associated with this feature — the change is in the file-discovery layer, not in the model or inference path.

What to watch after Claude Code now reads AGENTS.md if there

Teams that have intentionally kept Claude.md and AGENTS.md with different content in the same repo should audit their directory structure now. Under the new logic, AGENTS.md is silently ignored when Claude.md exists, so any Claude Code-specific instructions that were placed in AGENTS.md as a workaround will no longer be read. Those instructions need to move into Claude.md or the team risks running Claude Code with an incomplete context without any error message indicating the gap.

A second area to monitor is nested-directory behavior. Anthropic's changelog entry describes the fallback at the project root level; it does not specify whether Claude Code also walks subdirectories looking for AGENTS.md in the same way some tools support nested instruction scopes. Developers relying on per-directory AGENTS.md files in subdirectories should test that edge case explicitly before assuming the behavior extends beyond the root, and watch the Claude Code changelog for any follow-up entry that addresses directory traversal depth.