A blog post published on szypowi.cz reports a technical observation about Claude Code, Anthropic's command-line coding agent. According to the author, the AGENTS.md file — used to provide custom instructions to the agent — is only actually parsed and applied when telemetry, meaning the sending of usage data back to Anthropic, is enabled on the user's machine.
If accurate, this would point to a fairly common transparency issue in proprietary AI tooling: a feature advertised as available by default would in practice behave differently depending on settings that have no obvious logical connection to it. The fact that reading a local configuration file would depend on a remote data-collection setting raises questions about how these two mechanisms are technically coupled inside the product.
At this stage, the claim rests on a single observation shared via Hacker News, without official confirmation from Anthropic or a fully detailed methodology explaining how the behavior was tested. Developers who use Claude Code regularly may attempt to reproduce the finding, which would help clarify whether this is a bug, an undocumented dependency, or a misreading of how the tool actually works.
The report reflects a recurring tension around command-line AI agents, which are often treated as black boxes by the developers who integrate them into their workflows. How much real control users have over data sent to vendors, and which features implicitly depend on that data flow, remains a point of concern for technical teams deploying such agents in sensitive environments.