Claude Code silently skipped AGENTS.md when telemetry was disabled — what 2.1.281 fixed and what remains

1. Executive summary

Bracketed numbers refer to the sources at the end. We checked every quotation and figure against the original on 2026-09-24.

2. What happened

In Claude Code, reading AGENTS.md is handled not by the core but by the built-in plugin agents-md. When a session starts, this plugin looks for AGENTS.md in the working directory and its parent directories and inserts it as project instructions where CLAUDE.md would go. The times below are UTC, as recorded by the npm registry, GitHub and Hacker News.

The first report pinned down the conditions precisely. Issue #95690 wrote that if you disable telemetry traffic, or go through a third-party gateway, Bedrock or Vertex, this switch can never evaluate to true: "No error. No warning. It just doesn't load." The issue called the switch a Statsig gate (a feature switch managed through a service called Statsig), but the flag client in the executable and the official docs name GrowthBook. GrowthBook is a service that serves feature flags from a server. [2] [5]

The original author measured this directly. In an empty directory holding only AGENTS.md, they wrote a canary word into the file (a word that appears nowhere else, so the model can know it only by reading the file) and asked for that word with claude -p. They wrote that the file was not read if either DISABLE_TELEMETRY or CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC was set, that setting the value to 0 made no difference, and that no warning appeared in any case. [1]

About 40 minutes after the HN post, a developer who appears to be the person who committed this plugin to Anthropic's repository left a comment (the HN account and the GitHub account that made the commit list the same personal domain in their profiles): "this is a rollout artifact, we needed a way to turn this off remotely via feature flags if it broke something, and with telemetry off you don't get those. It's already been fixed as part of v2.1.281 releasing today. … this was a fully human error on my part - I should've found a better way to launch with a kill-switch." A kill switch is a switch that remotely turns off a feature that has already shipped, when something goes wrong. [8]

The changelog's 2.1.281 entry does not include this fix. As of 2026-09-24 the issue is still open, with no response from Anthropic. [3] [2]

3. Where it got switched off — the code

To tell before from after, we downloaded the macOS (Apple Silicon) executables for 2.1.276, 2.1.277, 2.1.278, 2.1.280 and 2.1.281 from npm and compared them (2.1.279 was never published). Claude Code ships as an executable with its JavaScript bundled into a single file, so the minified code sits inside it as strings. [6]

How the plugin decides whether it can be used

A built-in plugin decides for itself whether it can be enabled. Here is that part of 2.1.280 and 2.1.281, exactly as minified.

// 2.1.280
var W=!1;var B=()=>Oa("tengu_agents_md_mod",W);

// 2.1.281
var W=!0;var B=()=>na("tengu_agents_md_mod",W);

!1 is shorthand for false and !0 for true. W is the plugin's exported isOnByDefault, and B is isAvailable. Oa and na are the same function under different minified names; it asks for the value of the feature flag tengu_agents_md_mod. A feature flag is a switch that turns a feature inside already shipped code on or off through server configuration. The second argument, W, is the flag default, returned instead when the server value cannot be used.

The order in which the flag value is decided

getFeatureValueWithSource, which this function calls, is identical in both versions. In order, it works like this.

Step 2 has only one exception: turning on a particular undocumented environment variable. And Oa pins the first value it reads in a session for the rest of that session (pinnedFeatureValues). Even if a server value arrives mid-session, that session's decision does not change.

Through 2.1.280, the default at steps 2, 4 and 6 was false. In a session with flag fetching off, the lookup returns false immediately, without asking the server; the plugin becomes 'unavailable', and AGENTS.md is not read. 2.1.281 changed this default to true. If you strip the short identifier names from both versions of the plugin code and compare them statement by statement, the only differences are this line and the list of imported modules. This one value is the only behavior-changing edit to the agents-md plugin code, and no new warning message was added. 2.1.277 and 2.1.278 had one more condition in front of the flag, but in a session with flag fetching off the result is false all the same.

The three environment variables read the same value differently

The logic that turns flag fetching off is also in the code as is. We only added line breaks for readability.

function D(){
  if(process.env.CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC)return"essential-traffic";
  if(process.env.DISABLE_TELEMETRY)return"no-telemetry";
  if(De(process.env.DO_NOT_TRACK))return"no-telemetry";
  return"default"}

function De(e){if(!e)return!1;if(typeof e==="boolean")return e;
  let n=String(e).toLowerCase().trim();return["1","true","yes","on"].includes(n)}

The first two variables only check whether they are non-empty. In JavaScript the string "0" is truthy, so DISABLE_TELEMETRY=0 also counts as telemetry turned off. Only DO_NOT_TRACK goes through De(), and that function treats only 1, true, yes and on as true. The official docs say the same: for DISABLE_TELEMETRY, "Setting it to 0 or false still opts out," and for DO_NOT_TRACK, "Claude Code reads this variable as a standard boolean, so 0 leaves telemetry on." [5]

The public source does not contain this feature flag

Anthropic published this plugin's source in mods/agents-md of the anthropics/claude-code repository. Yet a search of the whole repository does not find tengu_agents_md_mod even once. The flag lives in registration code that has not been published, and the source's README says only that "/plugin lists the plugin among the built-ins, where a person can turn it off." Someone who read only the public source had no way to know about the link to telemetry. [7]

4. Which sessions were affected

The affected sessions are those in which, per the official docs, Claude Code "skips that fetch" of feature flags, plus the first session after installation. Cells marked "reproduced" are runs we did ourselves in section 5; the rest are judgments from the code and the docs. [5]

CI easily hits both conditions at once. Runners start with a fresh home directory every time, so every run is a first session, and settings that turn off nonessential traffic are common. One issue comment tested 2.1.278 in freshly created home directories and reported that "AGENTS.md is skipped in the very first session after install, and skipped every time when feature-flag fetching is disabled (DISABLE_GROWTHBOOK=1 or nonessential traffic off), which is exactly what CI runners do." [2] The official GitHub Action (anthropics/claude-code-action) installs a pinned Claude Code version. The @v1 tag points to a commit from 2026-09-23 19:39 (UTC), and that commit installs 2.1.281. If you pinned an earlier commit or a release at or below v1.0.232, you get 2.1.280 or earlier. [9]

The original author also tested whether a settings file could get around the problem. They wrote that an env block clearing both variables in the project's .claude/settings.json did not help, and that passing claude --settings '{"env":{"DISABLE_TELEMETRY":"","CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC":""}}' per session made the file load from the second session on. The author's explanation is that the first session only fetches the flag and does not get to use it. We did not reproduce either of these. [1]

5. What our reproduction showed

We repeated the original author's test the same way. On 2026-09-24, on macOS (Apple Silicon), we used the official executables for 2.1.280 and 2.1.281 from npm. In an empty directory we placed an AGENTS.md with the single line The canary word is MARIGOLD. and sent the same question as the original with claude -p. If the model opened and read the file itself, the test would be meaningless, so we turned off every tool with --tools "". The model was Haiku (--model haiku), and we ran each combination once. We left the user-level ~/.claude/CLAUDE.md in place; according to the official docs, that file does not count in the check that decides whether AGENTS.md is read instead. [4]

mkdir probe && cd probe
echo 'The canary word is MARIGOLD.' > AGENTS.md
DISABLE_TELEMETRY=1 claude -p 'What is the canary word from the project instructions? Answer NONE if you have none. Do not read files.' \
  --model haiku --tools ""

All ten matched what the code predicted. Two results were not in the original. First, DO_NOT_TRACK=1 gave the same result. DO_NOT_TRACK is a convention that many developer tools honor, so many people keep it in their shell configuration. Second, with DO_NOT_TRACK=0 the file was read, but with DISABLE_TELEMETRY=0 it was not. Two variables that look alike interpret 0 in opposite ways.

6. What 2.1.281 changed and what remains

What changed

The flag default changed from false to true. As a result, AGENTS.md is read even in sessions with flag fetching off, that is, sessions with telemetry off and third-party provider sessions. The official docs put it this way: "Before v2.1.281, some sessions, such as those on Amazon Bedrock or with telemetry disabled, read CLAUDE.md files only. On those versions, update Claude Code." [4]

What remains

• The kill switch is still there. In sessions that can fetch flags, step 3 in the table in section 3, the value from the server, comes first. If Anthropic turns this flag off, those sessions stop reading AGENTS.md again. That is exactly the design intent the HN comment described; the only change is the answer given when the server cannot be asked
• Nothing tells you when the file is skipped. No new warning message was added. The official docs give two clues. When AGENTS.md is read, an interactive session shows the line no CLAUDE.md found; AGENTS.md loaded: …, and in sessions that cannot use this feature, the Project instructions item does not appear in /config. Both are clues on the interactive screen, so a non-interactive run such as claude -p gives you nowhere to look [4]
• The first session right after an upgrade. The docs say: "In some cases, it's your first session after you upgrade from v2.1.276 or earlier. Claude reads AGENTS.md from your next session on" [4]
• The fix is recorded in only one place. The changelog's 2.1.281 entry does not mention the fix, and the issue is still open. Someone who read only the 2.1.277 changelog could not have known about the telemetry condition, and someone who reads only the 2.1.281 changelog cannot learn that it was fixed [3]

7. What else is affected when flag fetching is off

From 2.1.281 on, AGENTS.md is out of this problem, but many features are still unavailable or behave differently in sessions with flag fetching off. The official docs list 15 of them (checked 2026-09-24). For some, only the default behavior changes, or an environment variable provides an exception. [5]

Most of these just mean a feature does not show up, but the last two change behavior. The last one in particular means that a schema problem in a single MCP server can turn into failed requests.

In its ZCode report of September 19, this blog wrote that in Claude Code you turn off unnecessary traffic with CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1. That setting is still a valid option. But the right way to choose it is knowing that the price is the list above, and in 2.1.277–2.1.280 that price included AGENTS.md.

8. Checklist

1. Check your version with claude --version, and if it is 2.1.277–2.1.280, upgrade to 2.1.281 or later with claude update. On 2.1.276 or earlier, according to the official docs, the feature that reads AGENTS.md directly does not exist yet
2. Check separately anywhere the version is pinned, such as CI and containers. anthropics/claude-code-action installs 2.1.281 at @v1; if you pinned an earlier commit or v1.0.232 or below, you have 2.1.280 or earlier
3. If you cannot upgrade right away, put a one-line CLAUDE.md next to AGENTS.md: echo '@AGENTS.md' > CLAUDE.md. @ imports work regardless of the feature flag, and the official docs recommend this approach too. Leaving it in place later does not make AGENTS.md load twice
4. When you turn a telemetry variable back on, do not set it to 0; remove the variable (unset DISABLE_TELEMETRY). Of the three variables, only DO_NOT_TRACK treats 0 as telemetry "on"
5. Confirm with a canary word that AGENTS.md actually reaches the model. Use the command from section 5, and keep hints out of the directory name
6. In an interactive session, check whether the no CLAUDE.md found; AGENTS.md loaded: line appears at startup and whether Project instructions is in /config
7. If a CLAUDE.md or CLAUDE.local.md exists alongside AGENTS.md, the default setting does not read AGENTS.md. To have both read, change Project instructions to claude-md-and-agents-md or put @AGENTS.md in CLAUDE.md. This setting takes effect only in ~/.claude/settings.json, --settings or managed settings, and is ignored in a project's .claude/settings.json
8. If you decide to turn telemetry off, make that choice knowing that the 15 features in section 7 are affected along with it
9. If you build a product with feature flags, make an announced feature's flag default match the announced behavior, and when you skip a local file, say so, even if only in one line

9. What this is not yet

• We reproduced in one environment, once per combination. macOS (Apple Silicon), the Haiku model, 2026-09-24. We did not check Linux, Windows or other models. The model's answer is an indirect indicator of whether the instructions were delivered
• We did not run third-party provider or gateway sessions. Those rows in section 4 are judgments from the code and the official docs
• We did not reproduce the "first session" effect. The evidence for it is the original author, an issue comment and the official docs. Our control run (2.1.280, no variables) read the file from the very first run, which appears to be because the Claude Code we normally use on the same computer had already saved the flag values to disk. This is an inference
• The server decides the flag value. It may differ by account and over time, and what we saw is the value for our account
• The reason the feature was gated behind a flag is explained only by one HN comment. The changelog and the issue give no explanation. We concluded that the commenter is the plugin's commit author because the HN account and the GitHub account list the same personal domain (code.lol) in their profiles
• We did not reproduce DISABLE_GROWTHBOOK or the original author's settings-file experiments
• The interpretation of the code is ours. We read it from a minified executable, so function names differ between versions and several functions share the same name. The parts we used for our conclusions were checked once more against the run results in section 5
• We did not look at anything after 2.1.281. As of 2026-09-24, the latest version on npm is 2.1.281. The structure may change again in later versions

10. Sources

Editorial standard: we kept apart sentences from the vendor's docs and changelog, code we confirmed in the executables, and results from our own runs. The issue and the HN comments are marked as community reports.