Claude Code Fails to Load AGENTS.md Without Telemetry

Anthropic recently released Claude Code versions 2.1.277 and 2.1.280, introducing native support for project-level configuration via an AGENTS.md file. The update was designed to automatically detect and utilize the file when a standard CLAUDE.md configuration is absent, simplifying developer workflows. However, extensive testing and community reports indicate that the feature remains nonfunctional for developers who have disabled telemetry or operate through third-party infrastructure such as AWS Bedrock and Google Vertex AI. The limitation originates from the plugin architecture governing the feature. The agents-md loader depends on a remote feature flag designated tengu_agents_md_mod. When telemetry is disabled, Claude Code cannot reach Anthropic servers to resolve the flag. The system consequently defaults to an unavailable state, completely bypassing the local AGENTS.md file. Notably, the software issues no console warnings or diagnostic messages to indicate that the configuration file was skipped. Users are left unaware that their project instructions were never processed by the model, often prompting wasted effort on prompt engineering rather than identifying the underlying architectural blockage. This implementation disproportionately impacts privacy-conscious developers and enterprise environments that routinely restrict nonessential network traffic. Because the feature exclusively processes a local file, mandating a server-side approval gate has drawn sharp criticism. Industry observers and users argue that a transparent fallback mechanism or explicit error messaging would better align with modern software standards. Until Anthropic resolves the gating logic or introduces visible diagnostics, the established workaround requires creating a CLAUDE.md file containing an @path import directive that explicitly references AGENTS.md. This approach circumvents the remote flag dependency while preserving the intended single-file configuration structure. The incident highlights persistent friction between corporate feature rollout methodologies and developer demands for transparent, privacy-respecting tooling.