Claude Code's Background Agent Status Line Takes 16% to 26% of Usage in Three Users' Counts

Claude Code sends each background agent's full transcript to its model about every 30 seconds to produce a status line of three to five words, according to a gist published by Flamerix AI that was last active Oct. 29, 2026. The author's telemetry puts that traffic at 16.4% of spend.

A separate measurement on GitHub points the same way. In issue #86879 on the anthropics/claude-code repository, a user going by rflcrz counted 339 agent_summary calls in a 4-hour session with 27 subagents. They made up 18.0% of weighted token use. As of Oct. 29 the issue was open, with no assignee and no reply from Anthropic staff on the page.

What the call does

The gist says Claude Code asks the agent's model, usually Opus, to "describe your most recent action in 3-5 words". The result updates a status line in the agents panel and does not reach the conversation. The author says there is no setting to turn it off and that the behaviour became the default in version 2.1.232 in August 2026.

Because the request carries the transcript, the input is large and the output tiny. The gist reports an input-to-output token ratio of about 9,000 to 1. Issue #86879 reports 61.6M cache-read tokens against about 22,000 output tokens for the 339 summary calls.

Three measurements, three metrics

The figures are not interchangeable. Each user counted something different, and all of them are self-reported.

1. Flamerix AI, share of estimated spend16.4 %
2. rflcrz, share of weighted tokens18 %
3. sylvain-artois, share of cache-read tokens26 %

Source: Flamerix AI gist (estimated spend); GitHub issue 86879 by rflcrz (weighted tokens) and sylvain-artois (cache-read tokens), accessed 2026-10-29

Reporter	Claude Code version	What was counted
rflcrz	2.1.233	339 of 1,496 requests, 22.2% raw and 18.0% weighted
sylvain-artois	2.1.276	26% of cache-read tokens in one session
Flamerix AI	2.1.257 to 2.1.312	19.7% of API requests, 16.4% of estimated spend

The Flamerix AI gist says the spend figures are Claude Code's per-request estimates, not invoices, and that 96% of the summary spend went to Opus. Issue #86879 adds that the call is flagged skipTranscript: true, so tools that read transcripts to price a session will undercount it.

Workarounds and their costs

Setting CLAUDE_CODE_FORK_SUBAGENT to 0 in ~/.claude/settings.json stops the narration, a commenter named uwuclxdy reported in the issue, but it also turns off fork mode. The same thread says CLAUDE_CODE_ENABLE_NARRATION=0 worked on 2.1.241 and stopped working on 2.1.245.

The Flamerix AI gist offers a Bun preload script that replaces the summary timers with inert ones. It must be re-checked after every update, and the author says the agents panel stops updating. A commenter named JHVIW warned against it on Oct. 29 because it is fetched with curl, always pulls the latest version and runs inside every Claude Code process. The gist's own disclaimer says changing a client's runtime behaviour may conflict with the terms of the service.

What is not established

The sources are one gist and one issue thread, and the measurements come from the authors' own sessions, including one with 27 subagents. The issue asks for a supported setting, a configurable interval or coverage under CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC. A commenter on the thread found an undocumented CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES variable and asked maintainers what it does. Nobody had answered on the page as of Oct. 29.

The site has covered the review burden of agent-heavy work in Brooks's Law Meets AI Agents and agent tooling in Cursor's Agent SDK release. Neither measured background-agent overhead.

The number to watch is the issue's status: it has been open since Aug. 15 with the label area:cost.