On September 18, Claude Code v2.1.277 added support for AGENTS.md: in any project without a CLAUDE.md, Claude Code would read AGENTS.md instead — the emerging cross-tool convention for project instructions that Codex CLI, Cursor, and others also read. The changelog entry was honest about one limitation up front: “not yet on Bedrock, Vertex or Foundry.” What it didn’t disclose is the more interesting problem, surfaced two days later by a GitHub issue and independently corroborated by a Sept 23 deep-dive: reading a local markdown file off disk, in Claude Code, currently depends on a successful network call.
The bug#
AGENTS.md support ships as an internal plugin gated behind a Statsig remote feature flag, tengu_agents_md_mod, which defaults to off. Claude Code fetches that flag from Anthropic’s infrastructure at startup to decide whether to enable the loader. If the fetch fails — or never happens — the flag stays off and AGENTS.md is silently never read. No error. No warning. No line in /status. Your project instructions just aren’t there, and nothing tells you why.
That fetch fails, predictably, in exactly the environments where teams are most likely to have deliberately hardened their setup:
- Telemetry disabled. Set
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1— a documented, sanctioned way to reduce non-essential outbound calls — and the flag can no longer resolve. - Gateway/self-hosted deployments. Bedrock, Vertex, and third-party gateway setups (the exact audience Anthropic’s own Claude apps gateway targets) were already excluded per the changelog’s own disclosure — but the reason is the same remote-flag dependency, not a genuine platform limitation.
The reporter, GitHub user Prefee, filed the issue on September 20 with the label area:core and flagged it as a critical-impact “data loss or corrupted project” risk — not because files get deleted, but because an agent silently operating without the instructions you assumed it had can produce exactly that outcome. The framing in the issue title is blunt and correct: “AGENTS.md Support: A Local Feature Locked Behind a Remote Switch.”
Why gate a local file read at all#
There’s a legitimate reason Anthropic ships features behind Statsig flags: staged rollouts let a team catch a bad interaction before it hits 100% of users, and Claude Code has used this pattern for plenty of features that genuinely need it — anything touching model routing, auto-mode classifiers, or server-side behavior. The failure mode here isn’t “feature flags are bad.” It’s that a pure file-read — no model call, no API round-trip, nothing that could plausibly need a kill switch for safety reasons — inherited the same remote-gating pattern as features that actually depend on server state. The workaround the community found confirms this: dropping a one-line CLAUDE.md containing @AGENTS.md, using Claude Code’s local @path import syntax, bypasses the flag entirely because imports don’t route through the gated plugin. If the fallback path works with zero network dependency, the primary path didn’t need one either.
What to do about it today#
If your team is running Claude Code with telemetry disabled, through Bedrock/Vertex, or behind any gateway that might interfere with Statsig flag resolution, and you’re relying on AGENTS.md rather than CLAUDE.md for project instructions, do the workaround now: create a CLAUDE.md with a single line, @AGENTS.md, in every affected repo. It costs nothing and removes the dependency on a flag you can’t observe from inside the CLI.
As of v2.1.281, shipped September 23 — the same day as the Claude Marketplace launch — the changelog contains no entry referencing this issue, the Statsig flag, or a fix. GitHub issue #95690 remains open. Anthropic has shown a same-day patch cadence on higher-severity findings before (three security releases in three days back in early August), so a fix landing quickly wouldn’t be surprising — but as written today, this is still live.
The broader lesson#
This blog has spent all year making the case that Claude Code’s harness — not just the underlying model — is the real differentiator against IDE-anchored competitors. That argument gets weaker, not stronger, when the harness quietly reintroduces a network dependency into what should be a purely local operation, in exactly the enterprise deployment modes (self-hosted gateways, telemetry-restricted environments) that are supposed to be Claude Code’s strongest pitch to regulated organizations. The fix here is straightforward and the workaround is trivial. The lesson is less about this one flag and more about a habit worth checking for generally: when a CLI tool that runs on your machine claims a capability, it’s worth verifying that capability doesn’t have an invisible dependency on a server you don’t control.
Sources: anthropics/claude-code issue #95690 (filed Sept 20, 2026); Przemysław Szypowicz, “Claude Code reads AGENTS.md only when telemetry is on” (Sept 23, 2026); Claude Code changelog.