Claude Code reads AGENTS.md only when telemetry is on
First reported by Blog.szypowi.cz ·
Local instructions for code assistants now silently fail if telemetry is off.
Claude Code version 2.1.277 introduced support for AGENTS.md files, which provide project-specific instructions. However, this feature is gated behind a remote feature flag called 'tengu_agents_md_mod'. The loader for AGENTS.md is disabled by default and can only be enabled if this flag can be successfully fetched from Anthropic's servers. Consequently, if telemetry or non-essential traffic is turned off, or if the feature flag cannot be resolved, the AGENTS.md file is silently skipped without any warning to the user. This behavior was discovered and documented by a user who conducted tests by setting environment variables to disable telemetry and non-essential traffic, finding that the AGENTS.md file was not read. A workaround involves creating a CLAUDE.md file with a single line '@AGENTS.md' to manually load the instructions, which bypasses the remote feature flag.
The reliance on a remote feature flag for reading local project instructions like AGENTS.md means that users who prioritize privacy by disabling telemetry will be unable to use this intended functionality. This creates a direct conflict between a user's desire for privacy and the utility of the tool, particularly for developers who often manage sensitive codebases or operate in environments with strict data handling policies. The current implementation silently disqualifies a core group of users from a feature that is supposedly available.
This situation signals a broader tension in AI tooling between feature rollout strategies and user control over data. While staged rollouts and feature flags are standard engineering practices, their implementation should not come at the cost of silent feature failure, especially when the feature itself is local file access. The expectation for privacy-conscious users is that disabling telemetry should not affect local operations, and a lack of transparency here erodes trust and leads to wasted debugging effort.
AI-written summary. May contain errors.