Claude Code loads the AGENTS.md file only when telemetry collection is active. The behavior was documented in a technical write-up that reached the front page of Hacker News, where the thread collected 383 points and 217 comments.
The report identifies a conditional check in the client code. When the telemetry flag is disabled, the parser skips the AGENTS.md lookup entirely. Re-enabling telemetry restores the file read. The fix mentioned in the title appears to address the conditional, yet the original observation remains relevant for users who keep telemetry off by default.
How the Conditional Load Works
The client sends a telemetry heartbeat before scanning the project root for AGENTS.md. Without that heartbeat, the instruction loader never executes. This means custom agent rules—such as coding standards, tool restrictions, or repository-specific commands—stay invisible to the model.
Developers who disable telemetry for privacy reasons therefore lose the benefit of project-level instructions without any error message or log entry.
Community Metrics and Reactions
The Hacker News thread shows clear patterns in the 217 comments:
- Multiple users confirmed the same conditional behavior across different Claude Code versions.
- Several reports noted that AGENTS.md content reappears immediately after toggling telemetry back on.
- A subset of commenters described similar patterns in other Anthropic tools that gate context loading behind diagnostic flags.
Bottom line: The 383-point discussion indicates this is not an isolated edge case but a reproducible design choice.
How to Verify the Behavior
Users can reproduce the issue with these steps:
- Create an AGENTS.md file containing a distinctive instruction.
- Run Claude Code with telemetry disabled via environment variable or settings toggle.
- Observe whether the model acknowledges the instruction.
- Re-enable telemetry and repeat the query.
No additional tooling is required beyond the standard Claude Code binary and a text editor.
Privacy Trade-offs
Keeping telemetry off prevents Anthropic from collecting usage data. The current implementation forces a choice between privacy and functional project instructions. Teams that handle sensitive codebases often prioritize the former, which silently disables AGENTS.md support.
The conditional also raises questions about what other context files might be gated behind the same flag.
Comparison With Other Coding Agents
| Tool | Custom instruction file | Loads without telemetry | Notes |
|---|---|---|---|
| Claude Code | AGENTS.md | No | Conditional on heartbeat |
| Cursor | .cursorrules | Yes | Local file read |
| Continue.dev | config.json | Yes | Open-source loader |
| GitHub Copilot | .copilot-instructions.md | Yes | Workspace-level rules |
The table shows that competing tools separate instruction loading from diagnostic reporting. Claude Code remains the outlier on this dimension.
Who Should Pay Attention
Developers running Claude Code on air-gapped or privacy-restricted machines should test their current telemetry setting before relying on AGENTS.md. Teams that already keep telemetry enabled face no immediate change. Privacy-focused users evaluating multiple agents will find clearer behavior in Cursor or Continue.dev.
Practical Next Steps
Check the telemetry toggle in Claude Code settings and run a controlled test with a sample AGENTS.md file. If the conditional persists after the reported fix, file a follow-up issue with reproduction steps. For projects that must keep telemetry disabled, migrate instruction content to a tool that loads files unconditionally.
The episode underscores how small implementation details in agent clients can create unexpected privacy-functionality conflicts.
Top comments (0)