Claude Code users who keep telemetry disabled may need to verify that their project instructions are actually reaching the model. In a September 23, 2026 report, Przemysław Szypowicz says Claude Code 2.1.280 did not read AGENTS.md when either DISABLE_TELEMETRY=1 or CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 was set. His test used a canary word stored in AGENTS.md and asked a fresh Claude Code session to return it without reading other files.

Claude AI logo identifying the product discussed in the report
Claude AI logo; the article concerns Claude Code instruction loading.ja.wikipedia.org

Placed after the opening paragraph because the report concerns Claude Code and its AGENTS.md behavior.

The reported mechanism is more specific than a general telemetry failure. Szypowicz says the AGENTS.md loader is a built-in plugin whose default enabled state is false, and whose availability depends on a remote Statsig feature flag named tengu_agents_md_mod. If the flag cannot resolve to true, the plugin is unavailable and the local markdown file is never read. That means the limitation can affect a purely local operation—opening a file in the working directory—even though the file itself does not require network access.

The GitHub issue filed in the anthropics/claude-code repository independently records the same high-level concern: AGENTS.md support introduced in Claude Code 2.1.277 is described there as being behind an off-by-default remote gate, with telemetry-disabled configurations and some third-party routes unable to evaluate the flag. The issue is user-reported and remains a report rather than an Anthropic-confirmed diagnosis, but it is relevant supporting context because it is hosted in the product's public repository and describes the same observed behavior.

For readers managing multiple repositories, the practical question is whether a privacy setting can unexpectedly change which instructions the model sees. According to Szypowicz, the failure is silent: Claude Code starts and answers, but does not warn that AGENTS.md was skipped. That matters operationally because a user may spend time revising prompts or debugging model behavior when the missing context never entered the session. The supplied evidence does not establish how often this occurs, which Claude models are affected, or whether every deployment path behaves identically.

The report identifies one workaround: place a CLAUDE.md file next to AGENTS.md containing a single import line, @AGENTS.md. Szypowicz says this loads the same instructions even while CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 remains set, because CLAUDE.md imports do not depend on the reported feature flag. The tradeoff is that teams regain the extra file that AGENTS.md support was intended to avoid. The evidence also says clearing the variables through a project-level .claude/settings.json environment block did not help, while a session-level override worked from the second session onward.

The source packet does not contain an Anthropic statement confirming the implementation details or a fixed release, so the safest conclusion is narrower than the headline: in the tested Claude Code 2.1.280 setup, AGENTS.md loading appeared to depend on a remote gate that telemetry controls could prevent from resolving. Users who rely on AGENTS.md should test their own configuration with a harmless canary instruction, check whether a CLAUDE.md import changes the result, and avoid assuming that the presence of the file guarantees it was read.