What
A new startup notice, subagent-config-warning, can warn you when your setup encourages Claude to hand work off to subagents. Subagents are separate helper Claude instances that run part of a task.
- It reads your CLAUDE.md file, your installed skills and your custom agents.
- It sends that text to a small model, a classifier, which decides whether the config encourages delegating to subagents. The classifier runs under the source name
subagent_config_classifier. - If it does, a warning says that editing the config may save tokens on the current model.
- Results are cached in the global config. Classifier calls are limited to 10 per window of time.
- The warning is shown only a limited number of times, with a minimum number of days between showings.
- Three new per-user state keys keep track of this:
subagentConfigWarningShown,subagentConfigChecksandsubagentConfigClassifierCalls. - It is off by default through the remote config default
{ off: true }. - The environment variable
CLAUDE_CODE_SUBAGENT_CONFIG_WARNINGoverrides that default. A true-looking value turns it on with default settings, and a false-looking value turns it off.
The same release also adds three new usage-tracking events unrelated to this warning.
Why
When the warning is on, the text of your CLAUDE.md, skills and agents is sent to a model to be judged. The aim is to point out setups that spend extra tokens on subagents. Use CLAUDE_CODE_SUBAGENT_CONFIG_WARNING to force the warning on or off.