Flamerix AI: Claude Code agent summary burns 15-20% of tokens

Claude Code's agent summary works by sending each background agent's full transcript to its own model roughly every 30 seconds, asking for a 3-5 word description of its latest action. That answer serves only to refresh the status line in the agent panel, not to inform the model or the conversation, according to Flamerix AI's analysis. The feature has been enabled by default since Claude Code version 2.1.232, and Flamerix AI says it cannot be turned off through settings.

Flamerix AI's measurements, drawn from OpenTelemetry exports, put the feature at roughly 15-20% of overall token spending, with about one in five API requests tied to it. The share climbs to 18-26% when Opus models are used, making cost growth particularly noticeable there. To address the lack of an off switch, Flamerix AI built no-agent-summary.js, a Bun preload script of about 70 lines that wraps the global setTimeout() function and blocks callbacks containing a specific anchor such as "[AgentSummary] Timer fired", leaving other timers untouched. Because it loads via Bun's --preload option without modifying the Claude Code binary, updates do not disable it, though changes to the anchor would break it.

Flamerix AI lists the visible effect as zero agent summary requests and no per-agent panel status updates, with overhead rising from about 10ms to about 17ms as Function.prototype.toString is called per timer callback but cached in a WeakMap. Compatibility reports cover Linux, macOS and Windows, the last only limited. Separately, the Claude Code lab run by IT venture Arrowheads points to the anchor-change risk and notes that using the BUN_OPTIONS environment variable loads the script into every Bun on the machine, and says it does not recommend the workaround. A further issue reported on Opus concerns the User-Agent header including the user's email address.