Claude Code reads AGENTS.md only when telemetry is on [fixed]

Original Article ↗ Hacker News Discussion ↗

A bug in Anthropic’s Claude Code tool meant it only read the new `AGENTS.md` configuration file when telemetry was enabled, triggering concerns that basic local behavior was being gated behind data collection. The company later explained this as an unintended “rollout artifact” from using remote feature flags as a kill switch, and quickly shipped a fix, but many users criticized the design choice of tying file reading to remote flags and the growing complexity of the codebase. The episode fed broader worries about opaque, proprietary AI tooling that can remotely change behavior, the overuse of feature flags in client software, and the influence of AI-generated code and documentation on software quality.

Telemetry-gated AGENTS.md behavior

  • Users discovered Claude Code only read AGENTS.md when telemetry was enabled, despite release notes implying universal support.
  • Some reported it had behaved this way for months, blocking staged rollouts whenever telemetry was off.

Anthropic explanation and fix

  • A commenter from the team called it a “rollout artifact”: AGENTS.md support was behind a remote feature flag wired through the same path as telemetry.
  • With telemetry off, feature flags couldn’t be toggled, so AGENTS.md never activated.
  • They labeled it a human error, apologized, and said a new version removes the dependency.

Bug vs. intent debate

  • Some saw this as an obvious bug or rollout oversight.
  • Others suspected it was intentional to encourage turning telemetry on.
  • A few argued it was “working as designed” given the flag architecture, but that the design itself was flawed.

Design and architecture criticism

  • Many argued that reading a local file should never depend on remote state or telemetry, especially without any visible warning.
  • Several called the design “indefensible,” even if the specific behavior was accidental.
  • Others downplayed the impact, noting that non-telemetry users merely got the feature a bit later.

Feature flags, telemetry, and remote control

  • Some defended feature flags and kill-switches as standard for safe rollouts, especially at scale.
  • Others were uneasy that a desktop tool can have behavior remotely toggled, framing it as “remote control” of their machine.
  • Debate arose over whether feature flags must be coupled to telemetry or could be pull-only and privacy-preserving.

Mods/plugin system and complexity

  • AGENTS.md support was implemented via a new “Mods” extensibility system and function hooks.
  • Many saw this as massive overengineering for what appears externally like “read CLAUDE.md or, if absent, read AGENTS.md.”
  • Others countered that it’s a low-risk way to exercise a new plugin system.

Open vs. closed harness & configuration expectations

  • Confusion over GitHub repos: only mods/plugins are source-available; the core harness is proprietary.
  • Some criticized this as misleading and urged using open harnesses instead.
  • Several asked why filename handling isn’t simply configurable, or at least not hardwired to feature flags.

Workarounds and practical notes

  • Users described workarounds: symlinks, CLAUDE.md that just references @AGENTS.md, shared .agents/skills directories.
  • Note: by default, if CLAUDE.md exists, AGENTS.md is ignored unless a specific “claude-md-and-agents-md” setting is enabled.