When AI developer tooling attempts to orchestrate autonomous coding tasks, the invisible boundary between local filesystem execution and cloud telemetry dictates whether your infrastructure remains secure or quietly breaks. The recent rollout of Claude Code version 2.1.277 introduced support for AGENTS.md, a mechanism ostensibly designed to supply project instructions whenever a standard CLAUDE.md file is missing. Yet, a technical teardown of the version 2.1.280 bundle reveals a jarring design choice: this loader ships as an internal plugin called agents-md, which stubbornly stays disabled unless an external server flag validates its execution.
Because the plugin evaluates a remote flag designated as tengu_agents_md_mod—defaulting strictly to false—reading a local configuration file unexpectedly requires active external network connectivity. When your network policy suppresses those telemetry calls, the local instruction file fails to load entirely, leaving developers none the wiser.
"When Claude Code cannot fetch the flag, the plugin is unavailable, and the local file is never read."
In practical testing, enforcing security via environment variables such as CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 effectively bricks the feature. Similarly, setting DISABLE_TELEMETRY=1 shuts down local context injection without triggering a single console warning or error log. For an enterprise trying to maintain strict compliance, silent context stripping is a structural nightmare.
Failure Modes Across Cloud Gateways and Workarounds
For engineering teams operating inside hardened network perimeters, this flag-resolution architecture introduces persistent operational friction. Explicitly assigning a zero value to these configuration options offers zero relief, as setting either variable to 0 fails to override the default lock. Furthermore, defining an environment block within the project's .claude/settings.json that attempts to clear both variables has zero effect, rendering repository-level policy overrides completely useless.
This behavior extends directly to enterprise cloud deployments. Third-party gateways running on Amazon Bedrock or Google Cloud Vertex hit the exact same brick wall, as the flag cannot resolve to true in air-gapped or restricted proxy environments either. Developers maintaining cross-platform setups have fortunately identified structural workarounds: CLAUDE.md natively supports @path imports that bypass the flag entirely, allowing a single-line pointer to pull in AGENTS.md without requiring telemetry permissions.
How development teams will navigate these hidden telemetry dependencies as client-side agent configurations increasingly rely on cloud-toggled runtime flags remains an open question. Relying on remote switches to govern local file reads is not just fragile architecture—it is an unacceptable compliance risk for any organization that takes its source code security seriously.