Edited by Zeanan and Yang Wen
Is Anthropic going open?
Just now, Claude Code member Thariq Shihipar (@trq212) posted that Claude Code now supports AGENTS.md files.
Starting from version 2.1.277, if no CLAUDE.md file exists in the project directory, Claude Code will check for and use AGENTS.md instead. Users can toggle this behavior using /config.
This capability is built on Claude Code's mods mechanism. Anthropic has made agents-md a built-in mod and has open-sourced the related code.
Related source code link: https://github.com/anthropics/claude-code/tree/main/mods/agents-md
Thariq said that developers will also be able to build custom versions around project instructions in the future.
AGENTS.md is like a manual written by the project for AI agents, detailing the project's structure, how to install dependencies, which tests to run, coding standards, and more. The official AGENTS.md website explicitly refers to it as the "README for agents." This update by Claude Code is the first time this popular harness has been integrated into a cross-vendor standard for project instructions.
As soon as the news broke, netizens rejoiced, and the comment section was quickly flooded with messages like “Finally unified!” and “Thanks.”
Even Tibo, head of OpenAI Codex, came over to congratulate: “Great! That’s the spirit. Welcome to the light side.”
Keep in mind that AGENTS.md originally originated from OpenAI’s Codex. On the surface, Anthropic and OpenAI seem worlds apart, but now they’ve come together in the highly competitive Agent arena.
Claude Code now supports AGENTS.md
In default mode, Claude Code searches the project directory tree for AGENTS.md; if no project-level CLAUDE.md file exists along the project path, the found file is loaded as the project instruction.
When the Agent reads files in a project subdirectory, mod also supplements the corresponding AGENTS.md according to the directory hierarchy. If a CLAUDE.md file exists in the same directory, the existing CLAUDE.md rules are followed.
The relevant settings are located in the Project instructions under /config. Users can choose to use only CLAUDE.md, fall back to AGENTS.md if CLAUDE.md is unavailable, or load both file types simultaneously. The preferred approach can be adjusted based on the project's collaboration model.
This means Claude Code no longer needs to rely entirely on its own file naming conventions, significantly reducing the onboarding cost for teams already using AGENTS.md to manage project rules.
Shopify once brought up compatibility issues.
Earlier this year, Machine Heart reported that Shopify CEO Tobi Lütke considered banning Claude Code internally until it supported reading files such as AGENTS.md and .agents/skills.
When team members use Codex, Cursor, and Claude Code simultaneously, the same repository may have different configuration entry points. Developers using Codex read AGENTS.md, while those using Claude Code read CLAUDE.md. If the two files differ, each agent will receive different project rules.
These documents typically record the project structure, coding standards, testing requirements, development processes, and the skills that the Agent can invoke. They have become essential context for AI coding agents to understand the codebase.
Previous solutions provided by developers included creating symbolic links or adding @AGENTS.md to CLAUDE.md. These approaches are easier to implement in small projects, but in large monorepos, configuration files recursively apply throughout the directory tree, causing maintenance costs to rise rapidly.
Lütke believes that teams should not bear this type of synchronization work long-term. He refers to this additional burden as a "complexity tax."
Why does Anthropic insist on CLAUDE.md?
Anthropic previously also considered its own products.
Thariq explained that different model families have distinct behavioral characteristics, and the system prompt significantly influences model performance. Claude models have specific preferences for how skills, system prompts, and CLAUDE.md are organized, and Claude Code adjusts the system prompt accordingly for different models.
This approach echoes Anthropic’s discussions on context engineering. The project description, skills documentation, and system prompt together form the context required for the agent to complete its task.
This model-specific design helps optimize the Claude Code experience but also increases maintenance costs for cross-tool collaboration.
Supporting AGENTS.md is just the first step
This update first resolves the issue with the file read entry point; for teams already using AGENTS.md, Claude Code no longer requires additional symbolic links or reference configurations.
However, compatibility does not mean complete uniformity. By default, Claude Code will only fall back to reading AGENTS.md if there is no CLAUDE.md in the project path. If your team wants to load both files simultaneously, you’ll need to adjust the settings in /config.
In addition, .agents/skills and .claude/skills remain two separate directories. The team still needs to determine which rules belong to cross-tool common content and which are specific to Claude Code.
As AI coding agents become integrated into daily development workflows, project instruction files are evolving from tool preferences into codebase infrastructure. Claude Code’s inclusion in AGENTS.md reflects the growing reality that sharing project context across tools is now a necessary requirement for development teams.
Previously, Claude Code was one of the biggest absentees from this standard; with its addition, the trend of mainstream AI coding tools converging around "Agent README" has become much clearer. In the future, the cost of migrating between different Harnesses is expected to continue decreasing.
So the question is, which Harness is more user-friendly right now?
Reference link:
https://x.com/trq212/status/2101009392611278961?s=20
https://github.com/anthropics/claude-code/tree/main/mods/agents-md
© THE END
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by someone interested in AI.