Claude Code reads AGENTS.md only when telemetry is on

Claude Code’s new AGENTS.md support is effectively disabled when telemetry or nonessential traffic is turned off, or when a remote feature flag cannot be fetched, causing a local, disk-only capability to fail silently. For CIOs and technology leaders, this creates operational and governance risk: teams that rely on AI coding tools may see inconsistent behavior, wasted troubleshooting effort, and policy-driven environments—especially privacy-conscious or gateway-based deployments—missing documented features without warning. The broader implication is that IT organizations should treat AI developer tools like any other enterprise software dependency: validate feature behavior under security/privacy controls, require transparent failure modes, and avoid assuming that advertised local capabilities will work in restricted environments.