Claude Code AGENTS.md Support Fails Silently When Telemetry Is Disabled
Claude Code version 2.1.277 introduced support for root AGENTS.md files, but the loader depends on a remote feature flag that fails when telemetry is disabled. Sessions omit project instructions without throwing warnings on offline shells, Amazon Bedrock, or Google Vertex AI. Developers can bypass the flag using an explicit import in CLAUDE.md.
Impact: Medium
Why it matters
You can prevent Claude Code from ignoring your cross-agent project instructions by adding a one-line path import or cleaning session environment flags.
TL;DR
- 01Claude Code gates AGENTS.md behind the remote feature flag tengu_agents_md_mod.
- 02Disabling telemetry or running on Vertex AI / Bedrock causes instructions to silently fail without warnings.
- 03Add @AGENTS.md to CLAUDE.md to force deterministic local loading.
Key facts
- Introduced Version
- Claude Code 2.1.277
- Analyzed Bundle
- Claude Code 2.1.280
- Remote Flag Name
- tengu_agents_md_mod
- Fallback State
- false (disabled)
The Remote Feature Flag Trap
Claude Code 2.1.277 introduced support for loading AGENTS.md as standard project instructions when no CLAUDE.md exists. In version 2.1.280, the loader ships inside a bundled plugin called agents-md registered as:
var B = () => Oa("tengu_agents_md_mod", W);
The fallback value W is set to false. When network access to Anthropic feature flags is blocked, isAvailable evaluates to false, and local file reads are skipped entirely.
Impact on Offline and Enterprise Setups
The gate fails when either of the following environment variables is set:
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1DISABLE_TELEMETRY=1
Setting either variable to 0 does not re-enable the feature because the CLI treats any assigned value as true. Repository-level configuration in .claude/settings.json cannot override this behavior. Enterprise deployments running through custom gateways, Amazon Bedrock, or Google Vertex AI hit this failure consistently without any console warning.
Workarounds for Multi-Agent Repositories
To ensure instructions load regardless of network state, link the target file directly from CLAUDE.md:
echo '@AGENTS.md' > CLAUDE.mdClaude Code resolves @path imports locally without checking remote switches. For shared skills, Claude Code only reads .claude/skills rather than .agents/skills; creating a filesystem symlink from .claude/skills to ../.agents/skills keeps shared toolsets in sync.
Try it in 2 minutes
echo '@AGENTS.md' > CLAUDE.mdbash
✓ When to use
- Multi-agent codebases sharing instructions across Claude Code, Codex, and Cursor.
- Enterprise development environments operating behind air-gapped proxies, Amazon Bedrock, or Google Vertex AI.
✕ When NOT to use
- Repositories that already rely exclusively on a standalone CLAUDE.md file.
- Environments where standard Anthropic telemetry endpoints are reachable and telemetry is left enabled.
What to do today
- Check your environment for DISABLE_TELEMETRY or CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC flags.
- Verify instruction ingestion by querying a canary token with claude -p.
- Add an @AGENTS.md import inside CLAUDE.md to bypass remote configuration dependencies.
What the community says
“Anthropic performing arbitrary changes that are tangential but not specific to the premise of delivering a harness to package code files and prompts in context to their LLM breaks that trust for many.”
“Couldn't you just make a CLAUDE.md that says, "Read AGENTS.md in this same directory"?”
Sources