Flamerix AI says Claude Code's background agent summaries take 16% of its spend

According to Flamerix AI, Claude Code sends each background agent's entire transcript to the agent's own model, usually Opus, and asks it to describe the latest action in 3 to 5 words. The answer only updates the status line in the agents panel. The company says no setting turns this off, and the feature has been on by default for interactive background agents since version 2.1.232 in August 2026.

In the company's OpenTelemetry export, requests tagged agent_summary made up 16.4% of Claude Code spend, compared with 15.7% for the main conversation. Flamerix AI says about 98% of these calls read the transcript from the prompt cache, and that the figures are Claude Code's own per-request estimates rather than invoices.

The no-agent-summary.js hook loads through BUN_OPTIONS and wraps setTimeout and setInterval, returning an inert timer when a callback's source matches a summarizer anchor. It was tested on Linux across Claude Code 2.1.257 to 2.1.312. The authors warn that an update can silently bring the summaries back, and the side effect is that the agent panel status stops updating.

Developers running background agents on Claude Code may be spending a noticeable share of their rate-limit headroom on status line summaries rather than on their own work. Checking OpenTelemetry data for agent_summary requests is a direct way to measure this. The hook depends on matching source text, so it needs re-verification after every Claude Code update.