Claude Code Now Falls Back to AGENTS.md When CLAUDE.md Is Missing

Anthropic's Claude Code has started reading AGENTS.md files when a project has no CLAUDE.md, a small configuration change that signals a much larger shift in how AI coding agents are governed. The move, reported by The Register as Anthropic's decision to support OpenAI's markdown instruction specification, quietly folds Claude into a convention that competing tools have been building around for months. For developers, the practical effect is immediate: a single markdown file can now steer more than one agent.

A Shared File for a Crowded Toolbox

The logic behind AGENTS.md is straightforward. Rather than forcing every team to maintain a bespoke CLAUDE.md, .cursorrules, or Copilot instruction file, the community has been consolidating around one plain-text document that describes build steps, coding conventions, and architectural constraints. Claude Code now honours that file as a fallback, which means a repository can carry one authoritative set of instructions and let multiple assistants read it.

That fallback behaviour matters because CLAUDE.md retains priority. If a project ships both files, Claude still follows its native format first, and only reaches for AGENTS.md when the primary file is absent. The hierarchy preserves backwards compatibility for existing users while lowering the barrier for teams that want to standardise.

Why the Timing Is Not Accidental

The change lands as Anthropic's own tooling grows more autonomous. Reporting from Technology Org notes that Claude now leads roughly 26% of Anthropic's AI research work, a figure that underlines how much the company depends on its own agents. Meanwhile, commentary from XDA argues that developers should stop treating CLAUDE.md as a memory store now that Claude can genuinely remember context across sessions, freeing the file to serve as a specification rather than a scratchpad.

Competitive pressure is part of the picture too. Analysis from Buttondown contrasts Microsoft's Copilot, which hands users a visible cost dial, with Claude's approach of absorbing that complexity. Standardised instruction files reduce friction on both sides: they make agents cheaper to configure and easier to swap.

What Developers Should Do Next

For maintainers, the sensible step is to audit existing configuration files and decide whether AGENTS.md deserves to become the canonical source. Teams running multiple assistants can consolidate duplicated instructions into one document, then keep CLAUDE.md as a thin wrapper if needed. The convention is still young, but its direction is clear, and repositories that adopt it early will find it far easier to move between tools as the agent market keeps churning.