Anthropic adopts OpenAI's standards, and all agents will share the same AGENTS.md henceforth.

Edited by Zenan, Yang Wen

Is Anthropic about to go open?

Just now, Thariq Shihipar (@trq212), a member of the Claude Code team, posted that Claude Code has started to support the AGENTS.md file.

Starting from version 2.1.277, if there is no CLAUDE.md in the project directory, Claude Code will check and use AGENTS.md. Users can toggle this behavior via /config.

This capability is built on Claude Code's mods mechanism. Anthropic has made agents-md a built-in mod and released the relevant source code publicly.

Relevant source code link: https://github.com/anthropics/claude-code/tree/main/mods/agents-md

Thariq said that in the future, developers can also build custom versions around project instructions.

AGENTS.md is like an instruction manual written by the project for AI Agents, whose content aims to tell the Agent the project structure, how to install dependencies, which tests to run, what the code specifications are, and so on. The official AGENTS.md website also explicitly calls it "README for agents". This change in Claude Code marks the first time that this popular harness has joined the cross-vendor project instruction standards.

As soon as the news came out, netizens were delighted, and the comment section was quickly flooded with comments like "Finally unified" and "Thanks".

Even Tibo, the head of OpenAI Codex, came to congratulate: "Awesome! This is the right move. Welcome to the bright side."

It is worth noting that AGENTS.md originally came from OpenAI's Codex. Anthropic and OpenAI, which seemed to be at odds with each other on the surface, have now come together in the fiercely competitive Agent track.

Claude Code begins to support AGENTS.md

In the default mode, Claude Code will search for AGENTS.md along the project directory tree. As long as there is no project-level CLAUDE.md on the project path, the found file will be loaded as project instructions.

When the Agent reads files in subdirectories of the project, the mod will also supplement the corresponding AGENTS.md according to the directory hierarchy. If CLAUDE.md exists in the same directory, the original CLAUDE.md rules will be followed.

The relevant settings are located in Project instructions under /config. Users can choose to use only CLAUDE.md, fall back to AGENTS.md when there is no CLAUDE.md, or load both types of files at the same time. The specific method can be adjusted according to the collaboration mode of the project.

This means that Claude Code no longer has to fully rely on its own file naming conventions, and the access cost is significantly reduced for teams that already use AGENTS.md to manage project rules.

Shopify once brought the compatibility issue to the table

A while ago, Jizhizhixin reported that Tobi Lütke, CEO of Shopify, once considered disabling Claude Code internally in the company until it supported reading files such as AGENTS.md and .agents/skills.

When team members use Codex, Cursor and Claude Code at the same time, the same repository may have different configuration entrances. Developers using Codex read AGENTS.md, while developers using Claude Code read CLAUDE.md. If there are differences between the two files, different Agents will get different project rules.

These files usually record project structure, coding specifications, test requirements, development processes, and skills that Agents can call. They have become important context for AI coding Agents to understand the codebase.

The previous solutions given by developers include creating soft links, or writing @AGENTS.md in CLAUDE.md. Small projects can easily adopt these solutions, but for large monorepos, configuration files will take effect recursively along the directory tree, and the maintenance cost will rise rapidly.

Lütke believes that teams should not bear this kind of synchronization work for a long time. He calls this extra burden "complexity tax".

Why did Anthropic stick to CLAUDE.md?

Anthropic previously had its own product considerations.

Thariq once explained that different model families have different behavioral characteristics, and the system prompt will significantly affect model performance. Claude models have specific preferences for the organization of skills, system prompts and CLAUDE.md, and Claude Code will also adjust system prompts for different models.

This line of thinking echoes Anthropic's discussions on context engineering. Project descriptions, skill files, and system prompts together form the context required for Agents to complete tasks.

This model-specific design helps optimize the Claude Code experience, but it also increases the maintenance cost of cross-tool collaboration.

Supporting AGENTS.md is just the first step

This update first solves the problem of file reading entrance. For teams that already use AGENTS.md, Claude Code no longer requires additional soft links or reference configurations.

However, compatibility does not mean complete unification. In the default mode, Claude Code will fall back to reading AGENTS.md only when there is no CLAUDE.md on the project path. If the team wants to load both types of files at the same time, they need to adjust the settings in /config.

In addition, .agents/skills and .claude/skills are still two separate directories. Teams still need to decide which rules are common cross-tool content and which content only serves Claude Code.

As AI coding Agents are integrated into daily development workflows, project instruction files are evolving from tool preferences to codebase infrastructure. Claude Code's adoption of AGENTS.md indicates that cross-tool sharing of project context has become a real demand for development teams.

Previously, Claude Code was one of the biggest absentees from this standard. After Claude Code joined, the trend of mainstream AI Coding tools converging around "Agent README" has become much more obvious. It can be predicted that the migration cost between different harnesses will become lower and lower in the future.

So the question is, which harness is better to use now?

References:

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 Official Account JIQIZHIXIN (ID: almosthuman2014), authored by AI-focused contributors, and is authorized for release by 36Kr.