Claude Code's status summaries took 16.4% of one team's estimated spend

The few words describing a background agent’s progress can carry a surprisingly large cost in some Claude Code workflows, according to a new public discussion of measurements by Flamerix AI. The team says agent_summary accounted for 16.4% of its total estimated Claude Code spend over one measured week, and 19.7% of API requests. These are the team’s observations, not a universal share of every user’s tokens or bill.

The report reached Hacker News on October 29 at 03:33 UTC. Its gist was created October 26, with the cited revision dated October 28. The underlying complaint is older still: an August GitHub issue asks Anthropic for a supported way to disable the summaries. Today’s hook is renewed scrutiny and a published workaround, rather than discovery of a feature that launched today.

Estimated spend is not an invoice

Flamerix says it grouped Claude Code OpenTelemetry events by query_source and found repeated summary requests using a background agent’s transcript. The generated words update that agent’s progress description in the agents panel. This should not be confused with every piece of status text at the bottom of a terminal.

The report’s spend figure comes from Claude Code’s per-request cost_usd estimate, not an invoice. Cached input is cheaper than fresh input but can still contribute to the estimate when long transcripts are repeatedly read. The result therefore depends on context length, model choice, concurrent agents and workload. The older GitHub issue contains other users’ measurements with different denominators; those figures should not be combined into one general savings rate.

A workaround tied to implementation details

The MIT-licensed no-agent-summary.js hook loads through Bun’s BUN_OPTIONS=--preload mechanism. Its source wraps timer functions and checks callback text for summary-specific anchors, suppressing matching timers without editing the Claude Code binary. Flamerix reports Linux tests across versions 2.1.257 through 2.1.312 and says summary calls disappeared with suppression enabled. The visible tradeoff is that generated per-agent progress summaries stop updating.

The documented setup requires placing the script locally, passing its absolute path into the Claude Code process environment, and restarting the daemon and existing sessions so they inherit it. It uses the user’s existing Claude Code access; it is neither an Anthropic setting nor a separate model provider.

macOS and Windows are untested in the report. Because matching depends on internal callback text, an update can silently restore the requests. The author recommends checking telemetry after updates. AITrending has read the report, hook and public discussion but has not installed the script or independently reproduced its savings. The evidence supports a concrete workflow-cost dispute and an inspectable experiment, not a guaranteed 16.4% reduction for all users.