Claude Code reads AGENTS.md only when telemetry enabled

The practice of gating software functionality behind remote feature flags is a standard deployment strategy in SaaS and developer tooling, allowing vendors to incrementally roll out capabilities while monitoring for regressions. However, when these flags are cryptographically or logically bound to telemetry pipelines, they create a class of 'privacy-tax' bugs where opting out of data collection results in functional degradation. This pattern has precedent in previous controversies involving IDE extensions and cloud-native CLIs where offline modes were broken by mandatory license checks disguised as analytics. In the AI safety domain, this intersects with 'tiered trust' architectures, where vendors distinguish between general consumers and vetted researchers. Safety controls often rely on server-side inference or logging; when local clients cannot verify their tier status due to disabled telemetry, they typically fail closed to preserve safety guarantees. This architectural assumption prioritizes risk mitigation over local autonomy, creating recurring friction points for security-sensitive deployments that require both advanced features and network isolation.

On September 23, 2026, a technical controversy emerged regarding Anthropic’s Claude Code command-line interface after user pszypowicz published findings alleging that the tool fails to load local configuration files when telemetry is disabled. According to a blog post hosted on szypowi.cz and discussed on Hacker News, the specific file in question, AGENTS.md, is intended to provide project-specific instructions and context to the AI agent during coding tasks [1]. The allegation asserts that this functionality is not universally available but is instead gated behind a remote feature flag that requires active telemetry or non-essential network traffic to resolve [7]. When a user disables telemetry, the application reportedly cannot reach Anthropic’s servers to check the status of this flag, causing the system to default to an unavailable state and ignore the local AGENTS.md file entirely [8].

This dependency creates a functional paradox for privacy-conscious users and security-hardened environments. According to commentary on Hacker News, Anthropic commonly places new features behind feature flags that require telemetry until those features are fully promoted and defaulted on for all users [6]. Critics argue that this architecture effectively forces users to choose between maintaining strict telemetry hygiene and accessing core configuration capabilities necessary for complex agentic workflows. The claim suggests that the inability to read AGENTS.md without telemetry is not merely a bug but a structural consequence of how feature rollout and telemetry infrastructure are coupled within the Claude Code client [7].

The timing of this allegation coincides with broader discussions about Anthropic’s safety architecture and its bifurcation between general users and elite researchers. While Anthropic has not yet responded to the specific allegations regarding telemetry-dependent configuration parsing, recent reports indicate the company is shifting toward a post-hoc monitoring regime for vetted biology labs, allowing them to disable real-time safety refusals for up to six months per project [4]. This policy shift, which replaces real-time blocking with 30-day after-the-fact monitoring, highlights a growing divergence in how safety and utility are balanced across different user tiers [4]. The AGENTS.md issue may be interpreted as a downstream symptom of this tiered architecture, where advanced configuration and relaxed safety constraints are increasingly tied to server-side validation mechanisms that assume persistent connectivity.

Furthermore, the controversy intersects with high-stakes scientific discovery efforts utilizing Claude agents. Recent disclosures describe instances where hundreds of Claude agents screened vast biological datasets over extended periods, identifying novel enzyme systems such as ART, a bacteriophage enzyme pairing a reverse transcriptase with a CRISPR-like repeat array [2][3]. These autonomous research tasks rely heavily on precise agent configuration and sustained operational stability. If local configuration files like AGENTS.md are inaccessible without telemetry, researchers operating in air-gapped or restricted-network environments may face significant friction. Although Anthropic states it does not yet know the function of the discovered ART system and notes that related reports have not been peer-reviewed, the reliance on autonomous agents for such discovery underscores the criticality of reliable configuration loading [3].

The technical dispute also raises questions about the transparency of feature gating in AI development tools. According to analysis from OpenX, the loader for AGENTS.md support was added in version 2.1.277 but remains situated behind a remote feature flag [7]. This implementation detail means that even though the code to parse the file exists locally, the permission to execute it is contingent on external server responses. For users who have opted out of telemetry, this results in a silent failure where the agent proceeds without crucial context, potentially leading to hallucinated workflows or non-compliant code generation. The critic’s position emphasizes that this coupling undermines user agency, as the opt-out mechanism for data collection inadvertently degrades product functionality [1].

As of the current timeline, Anthropic has maintained silence on the specific telemetry dependency claim. The absence of an official response leaves unresolved whether this behavior is an intentional design choice to enforce telemetry participation for beta features or an unintended side effect of the feature flag infrastructure. The situation reflects a tension common in modern software distribution where cloud-native management strategies collide with local-first privacy expectations. Until Anthropic clarifies the intended behavior or decouples the feature flag resolution from telemetry endpoints, users must navigate a trade-off between privacy and full access to Claude Code’s agentic configuration features.

Developers and researchers operating in air-gapped or privacy-restricted environments face degraded Claude Code functionality, losing access to AGENTS.md configuration essential for complex agentic workflows. Anthropic risks alienating security-conscious enterprise and academic users who view telemetry coupling as coercive. Concurrently, the company is expanding high-trust access for elite biology labs with 6-month safety exemptions, creating a two-tier system where verified researchers gain more freedom while privacy-focused general users lose basic utility. The magnitude affects the subset of Claude Code users who disable telemetry, but the reputational risk extends to all users evaluating Anthropic’s commitment to local-first operation. If unresolved, this could push security-sensitive organizations toward competitors offering true offline parity.