Claude Code's background agents spent more on a three-word status line than on the main conversation

Flamerix AI published a technical write-up on GitHub this week. It shows that Claude Code, Anthropic's coding agent, spends a large share of its users' tokens on a feature most of them never switched on. Roughly every 30 seconds, for every background agent, Claude Code sends that agent's entire transcript to the agent's own model, usually Opus. It asks the model to describe its most recent action in three to five words. The answer updates a status line in the agents panel and goes nowhere else. In the week before Flamerix blocked it, this call, logged as agent_summary, took 16.4% of the team's Claude Code spend and 19.7% of its API requests, about one in five. The main conversation took 15.7%. According to the write-up, no setting turns it off, and it has been on by default since version 2.1.232 in August.

The cost sits where nobody looks

Read the figures with Flamerix's own caveats. "Spend" means Claude Code's per-request estimate, cost_usd, and is not an invoice. On a subscription, what you lose is rate-limit headroom. Of that spend, 96% ran on Opus, at about 9,000 input tokens for every output token. Flamerix also points to a GitHub issue, #86879, that collects four independent measurements of 18–26%.

The detail that matters most for anyone running agents is where the cost shows up. Claude Code does not write these requests to the session logs on disk. Only the OpenTelemetry export shows them, so cost tools that read transcripts miss them entirely. DRIM detected the anomaly in that export by grouping claude_code.api_request events by query_source. According to the write-up, the feature's weekly share of spend grew from 3.5–7% in July to about 16% by mid-October.

The controls belong to the vendor. The server sets the interval, clamped between 30 and 240 seconds, and the default is already the minimum. Version 2.1.310 ships a cheaper short mode that uses about 2k tokens of notes in place of the full transcript. It is switched off, and only server flags can turn it on.

AI does the routine work. You do the thinking. Part of that thinking is checking what the machine spends while it works. The person who delegates to agents and never reads the meter has handed part of their budget to someone else's defaults.

The fix, and its terms

Flamerix released an MIT-licensed Bun preload of about 80 lines. It wraps the runtime's timers and neutralises the one that fires the summary. Running it under every Claude Code process, the team reports zero agent_summary requests, and no other side query grew to take its place. Subagents and background sessions run normally. The visible cost is that the per-agent status in the panel stops updating.

The workaround has limits, and Flamerix lists them. An update can silently undo it, so you have to re-verify after every release. It has been tested on Linux only. Flamerix also warns that changing a client's runtime behaviour may conflict with the service's terms. Two environment variables, CLAUDE_CODE_FORK_SUBAGENT=0 and CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1, also stop the summaries, but they switch off fork subagents or background agents along with them.

Your Next Move

1. Read your own meter this week. Turn on Claude Code's OpenTelemetry export, group claude_code.api_request events by query_source for seven days, and put agent_summary next to the main conversation. Ask the same question of every agent tool you run: what share of spend reaches the work?
2. Take the numbers to whoever owns the account before you patch anything. On a team or company plan, the decision about running a runtime hook against a vendor's client belongs to whoever answers for the terms. If they approve it, add a verification step to every update.
3. Report cost per outcome. If your employer counts AI usage in reviews, tokens spent on a status line raise your usage figure without adding any impact. Our guide covers how to show the impact of your AI use before your next review. Spend data like this is the evidence that makes the case.