Anthropic Adds AGENTS.md Support, Strengthens Claude Code Compatibility

On September 18, 2026, Anthropic announced support for AGENTS.md, the instruction-file specification led by OpenAI, in its AI coding assistant Claude Code. It applies to version 2.1.277 and later, released the same day. Under the new behavior, if no CLAUDE.md exists in the same location, Claude will look for and use AGENTS.md. According to reporting by Thomas Claburn of The Register, the move is seen as reducing the compatibility burden for the developer community.

This update is notable because Anthropic has accepted a specification created by a competitor. Until now, the company had strongly favored its own format. Even so, this can be seen as a decision to prioritize interoperability.

Claude Code Adds Support for AGENTS.md

Claude Code engineer Thariq Shihipar announced the support in a post on social media. According to his explanation, users can toggle the behavior with a settings command. His exact statement is as follows.

We’re adding support for AGENTS.md to Claude Code. Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md.

With this change, if CLAUDE.md exists, it will continue to take precedence as before. Only when it does not exist will AGENTS.md be used as a substitute. By clarifying the priority, the change avoids unexpectedly altering behavior for existing users. A toggle via settings commands is also provided, allowing organizations to maintain their own operational policies.

Thibault Sottiaux, a core product lead at OpenAI, welcomed the announcement. He posted, “Yay! This is the way. Come to the light.” Shihipar responded with a handshake emoji. The public show of agreement between engineers at the rival companies can be seen as a symbol of specification convergence.

The Burden of Dual Maintenance and What

Compatibility Means

The instruction files referenced by coding assistants are context loaded for each request. They describe expected behavior, tools to use, coding conventions, verification procedures, and more. They form the foundation of the tool’s context and are directly tied to quality and consistency.

The challenge had become apparent in workplaces using both Claude Code and OpenAI Codex and similar tools. That is because teams had to maintain two separate files, CLAUDE.md and AGENTS.md. With additional per-project and per-directory variants, the number of files to manage kept growing. The two formats are similar but not identical. CLAUDE.md can include Claude-specific instructions. AGENTS.md, by contrast, aims to be a tool-agnostic common format.

In practice, many teams used symbolic links to keep the two in sync as a workaround. Every change required copying or updating links, creating operational friction. The new support is an attempt to absorb that friction on the tool side. As shown in LangGraph vs AutoGen vs CrewAI: AI Agent Comparison 2026, the more tool options there are, the more important a shared context becomes. The ability to reuse a common file across multiple tools is highly significant.

From Standards Competition to a Turning Point

Toward Cooperation

AGENTS.md is a specification proposed by OpenAI. Last year, the company donated the specification to the Agentic AI Foundation under the Linux Foundation. The aim was to move it out of single-company control and develop it as an open foundation. As of December 2025, more than 60,000 public projects had adopted AGENTS.md. That adoption cemented its status as the de facto common format.

The widespread use of Claude Code had given Anthropic room to maintain its own specification. Still, its shift toward cooperation can be seen as reflecting changes in how the tools are used. Development setups using multiple tools together have become common. There was little benefit left in keeping specifications divided. User demand likely pushed specification convergence forward.

Recent relations between the two companies show more than just rivalry. In Anthropic and OpenAI Propose Embedded External Evaluators, cooperation was discussed within a safety-evaluation framework. Interoperability for this file specification can be seen as a move in the same direction. A stance of balancing competition and cooperation is taking hold.

From a safety perspective, how instruction files are handled is becoming increasingly important. As shown in Anthropic to Release Highly Risky AI “Claude Mythos” on a Limited Basis, the more powerful the tool, the more its context management matters. Widespread adoption of a common specification provides a basis for unifying audit and verification procedures. In that sense, this update is not just about convenience.

Impact on Development Workflows and Remaining

Challenges

The direct benefit is freedom from maintaining two files. Shared content can be consolidated in AGENTS.md. CLAUDE.md need only be added when Claude-specific adjustments are required. Preparing templates for new projects and migration work will also be simplified. For small teams and individual developers, the reduced burden is clear.

At the same time, understanding the priority order is essential. If CLAUDE.md exists, AGENTS.md will not be referenced. If the two contain conflicting instructions, the unintended one could take precedence. Organizations need to define in their operating rules which file is the canonical source. They should also reaffirm policies against including confidential information in instruction files. The wider the sharing enabled by commonization, the more important information management becomes.

Tool-specific extended syntax and differences in interpretation may remain. Even under the same specification name, behavior may not match completely across tools. Teams need a process to verify actual loading results during testing. Commonization is only a starting point; optimizing operations remains the role of teams on the ground.

Editorial Opinion

In terms of short-term impact, we expect templates and file operations to converge over the next three to six months. Teams using multiple tools will likely make AGENTS.md the canonical source. Migration support and converter tooling will also accelerate. We assess that the toggle via settings commands will support phased adoption.

From a long-term perspective, we expect convergence on the common specification to progress over one to three years. Discussions on extensions and version management centered on the Agentic AI Foundation will likely intensify. Audit and compliance procedures may also become standardized. We expect differences between tools to be consolidated into tool-specific supplemental files.

As questions from the editorial team, clarification of priority and responsibility remains an issue. When CLAUDE.md and AGENTS.md coexist, which should be the canonical source? Where is the boundary for what information may be included in a shared file? The impact of expanded interoperability on safety evaluations will also need examination.

References

• “Anthropic decides to support OpenAI’s markdown instructions spec”, by Thomas Claburn — The Register, 2026-09-18T21:33:16.000Z (ARR)
• Source URL: https://www.theregister.com/ai-and-ml/2026/09/18/anthropic-decides-to-support-openais-markdown-instructions-spec/5297588