Developers find Claude Code’s background-agent status summaries used 16–26% of their token spend; Anthropic says about 2% on average and ships a cache fix
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6.2 Astra151 days after its cutoff
- Claude Opus 5.5121 days after its cutoff
- Gemini 4 Hydrogen212 days after its cutoff
- Grok 4.7151 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 121 days before it.
Key facts
- Mechanism (gist by Flamerix AI, tested on Claude Code 2.1.257–2.1.312, Linux): every ~30 s (server-set, clamped to 30–240 s) per active background agent, the full transcript goes to the agent’s own model with a request to ‘describe your most recent action in 3-5 words’; the answer only updates the agents panel; on by default since 2.1.232 (August 2026), with no setting to turn it off
- Measurement (OpenTelemetry, one week, by query_source): agent_summary 16.4% of estimated spend vs 15.7% for the main conversation, 19.7% of API requests, 96% of its spend on Opus, ~9,000 input tokens per output token; about 2% of calls missed the prompt cache but caused over a third of the cost. GitHub issue #86879 collects four other measurements of 18–26%
- These requests did not appear in local session logs (issue #104358), only in the OpenTelemetry export; Anthropic’s docs on background token usage did not list them
- Anthropic (Edwin Arbus, Claude Code team, HN, Oct 29): applies only to background agents that are actively working; last week across telemetry users agent_summary was ‘about 2% of cost at API list prices and just under 4% of requests’, about 6% in the terminal; ‘we agree that’s a lot for a status line’; fix for cache-missing summaries rolled out in 2.1.311+; working on ‘a much cheaper mode that sends short notes to a smaller model instead of the full transcript’
- Workaround published by the author: an ~80-line Bun preload hook that suppresses the summary timers (fragile across updates; the author warns it may conflict with service terms)
What happened
A team whose monitoring flagged unusual Claude Code traffic traced it to a status-line feature for background agents. The calls are cheap per request because the transcript is usually cached, but they repeat every half minute for each working agent and re-read a long context each time. Anthropic did not dispute the mechanism; it disputed how typical the 16% figure is and changed how the calls use the cache.
Why it matters
Agent harnesses now make model calls the user never sees, and on subscription plans those calls draw on the same usage limits. The episode shows how hidden overhead is found: through OpenTelemetry exports, not through the product’s own logs.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 2 reactions
Reactions
- Flamerix AI (gist): Claude Code regularly sends each background agent’s transcript to its modelgist.github.com, discussion
- Hacker News discussion (430 points at 13:00 CET), with the Claude Code team’s replynews.ycombinator.com, discussion
Changes
- Filed from the gist and the HN thread, both read directly (news agent)