Claude Code only reads AGENTS.md when its telemetry is switched on, according to a bug report filed against Anthropic's coding agent on September 20, 2026. The feature arrived in version 2.1.277. Anyone who has turned off usage reporting, or who runs the tool through Amazon Bedrock or Google Vertex, gets no AGENTS.md at all. Nothing on screen says so.
The report is issue #95690 in the public Claude Code repository. A blog post published on September 23 walks through the same behavior and a workaround.
What AGENTS.md support was supposed to do
AGENTS.md is a plain text instruction file that several coding agents share. It tells an agent how a project builds, tests and is laid out. Claude Code has its own file for that job, called CLAUDE.md.
Version 2.1.277 promised a simple rule. In the words the issue quotes from the release notes, "in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead." Tech AI Wire covered that change when it shipped: Claude Code 2.1.277 reads AGENTS.md as a fallback.
Why it stops working without telemetry
The issue says the feature is not built into the normal loader. It is a built-in plugin that is off by default. A remote switch turns it on.
That switch is a feature flag named tengu_agents_md_mod, served by Statsig. Statsig is an outside service that lets a company turn features on or off for users without shipping a new version. To check the flag, Claude Code has to contact that service.
It cannot do that in several common setups, the issue and the blog post agree.
| Setup | AGENTS.md loaded? |
|---|---|
| Default install, telemetry on | Yes, once the flag is on |
DISABLE_TELEMETRY=1 set |
No |
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 set |
No |
| Through Bedrock or Vertex | No |
In each "no" case the file is simply skipped. The issue describes the result bluntly: "No error. No warning. It just doesn't load." The reporter marked the impact as critical.
The complaint: a local file behind a remote switch
Both sources make the same argument. Reading a Markdown file from the project folder needs no network. Tying it to a server-side flag means a privacy choice quietly disables an unrelated feature.
The issue calls it "reading a local markdown file, an act that requires no network whatsoever," that "is silently locked away by a server-side switch." The blog post puts it as a rule: "A privacy setting should never quietly switch off unrelated local behavior."
The workaround
The blog post gives a fix that works today. Create a CLAUDE.md in the project and have it import the other file with one line: @AGENTS.md. Claude Code always reads CLAUDE.md, and the import pulls in AGENTS.md along with it. No flag is involved.
What this means for developers
Check whether your setup falls into one of the "no" rows above. Teams that disable telemetry for company policy, and teams that reach Claude through a cloud provider, are the likely cases. If either applies, assume AGENTS.md is not being read, whatever the release notes say.
The one-line CLAUDE.md import is the safe default for now. It costs a single small file and behaves the same on every setup. If you had deleted a CLAUDE.md after the 2.1.277 release, put a thin one back.
Test instructions rather than trusting them. Ask the agent to repeat a rule that exists only in AGENTS.md, such as your test command. If it cannot, the file was not loaded. Do this once per environment your team uses.
Watch issue #95690 for a fix. The broader lesson is that a tool's behavior can depend on switches you cannot see. Keeping instructions in a file the tool always reads is the dependable choice.
This article was first published on Tech AI Wire.
Also available in
Deutsch · 日本語 · Français · Español · Português
Top comments (0)