Claude Code had a silent telemetry-gated bug

A reported bug in Claude Code meant it only read AGENTS.md files when telemetry was switched on. That behaviour has now been fixed, but the issue struck a nerve, drawing 399 points and 228 comments on Hacker News. You can read the details in the original report.

For a business building agentic workflows, this is not just a developer gripe. AGENTS.md is part of an emerging convention for giving coding agents project-specific instructions. If an agent ignores those instructions based on an unrelated setting like telemetry, teams can get inconsistent results without knowing why. One developer may see the agent follow the rules. Another may see it miss them entirely.

The bigger issue is trust. Agent workflows depend on clear assumptions about what the tool reads, remembers and follows. When those assumptions are hidden or behave differently across settings, your process becomes harder to test and govern. The fix matters, but it is also a reminder to test AI workflows under the actual configurations your team uses, including privacy and telemetry settings.

This is the kind of thing we build into an AI command centre, where teams can track how their AI processes are configured and where failures might affect real work.