Anthropic’s Claude Code changelog for version 2.1.277 (September 18, 2026) adds a line that developers had been asking for: “Added AGENTS.md support - in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead”, configurable under Project instructions in /config, and not yet available on Bedrock, Vertex or Foundry. The Hacker News thread on the change drew hundreds of comments within a day, mostly from people relieved that one shared convention is winning.

Key facts

Fact Detail
Question answered Claude Code Now Reads AGENTS.md
Short answer Claude Code 2.1.277 added AGENTS.md support: with no CLAUDE.md in the project, it reads AGENTS.md and lets you change it under Project instructions.
Reading time ~2 min
Last updated 2026-09-20

Why one file matters more than it sounds

AGENTS.md is a plain Markdown file at the root of a project that tells coding agents how to behave there: commands to run, conventions to follow, things to avoid. The complaint that drove it was simple - every client wanted its own file, so a team maintaining project rules ended up writing them three times.

Claude Code adopting the same file means the rules you already wrote keep working. That is a small changelog line and a large practical difference for anyone who maintains a repository that several agents touch.

The nuance worth reading twice

  • CLAUDE.md still wins. AGENTS.md is the fallback read when no CLAUDE.md exists. If you keep both, you have two sources of truth - make one authoritative and have the other link to it.
  • Not everywhere yet. The changelog is explicit that Bedrock, Vertex and Foundry users do not get it yet, so a team on those platforms should not plan around the fallback.
  • It is a contract, not a personality. The file is most useful when it states verifiable rules - “run npm test before proposing a change” - rather than vibes.

What to put in your own instruction file

If you run an agent that touches your own machine rather than a repository, the instruction file is still the right place for the boundaries:

  1. Scope - which folders the agent may read and write. Naming the excluded ones matters more than naming the included ones.
  2. Hard stops - the actions that must always be approved: deleting, moving outside the project, sending anything, spending anything.
  3. Reporting - what “done” looks like, and where the record of what changed should live.

The file sets the expectation; it is not enforcement. On a desktop agent the enforcement is the approval step plus the action log: the instruction says what should happen, the gate makes sure it does.

MOOGH follows that split on Windows - the task plan is shown before anything runs, execution waits for your approval, and the history stays readable afterwards. Download MOOGH if you want the file to describe the rules and something else to enforce them.

Related reading: