Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Measure what Claude Code's background-agent summaries cost you, and decide what to turn off

Written November 1, 2026

For a Claude Code user running background agents: the short status line under each working agent is written by a model call that rereads the agent's full transcript, about every 30 seconds. Those calls do not appear in the session logs on disk. Measure your own share before changing anything, because the share depends on how you work.

What the summary call is

When a background agent is actively working, Claude Code sends that agent's entire transcript to the agent's own model, usually the strongest model in use, and asks for a description of its most recent action in three to five words. The answer updates only the status line in the agents panel. It never reaches the model doing the work or your conversation. Flamerix AI's write-up dates the behavior to version 2.1.232, in August 2026, when it became on by default for interactive background agents, and reports that there is no setting that stops these calls while leaving background agents running. Idle background agents do not get a summary call, according to a member of the Claude Code team writing on Hacker News on 29 October 2026.

The cost shape is lopsided. In the one team Flamerix AI measured, the summary calls carried roughly 9,000 input tokens for every output token, because a long transcript goes in and a handful of words comes out. Most of those calls reread the transcript from the prompt cache, which keeps each one cheap. The small share that misses the cache carries a large share of the cost: Flamerix AI puts it at over a third of the summary cost from about 2% of the calls.

Two populations, two very different shares

Flamerix AI measured one team's week and found the summary calls at 16.4% of that team's estimated Claude Code spend, against 15.7% for the main conversation, and 19.7% of all API requests, with 96% of the summary spend on the strongest model. That was one team, in one week, running mostly background agents. It is not a figure for a typical user.

Anthropic's own reply, from the same Hacker News thread, gives the wider population: across Claude Code users who send telemetry, the summary calls were about 2% of cost at API list prices in the previous week, and just under 4% of requests, rising to about 6% of cost in the terminal, where nearly all of these calls happen. The same reply says the 16.4% week sounds like a team running mostly background agents and does not seem like a typical session. Both figures can be true. They describe different workloads. Your share is the only one that should drive a decision, and the next section shows how to get it.

Measure your own share

Claude Code's OpenTelemetry export is the route Flamerix AI used, because the summary requests are not written to the session logs on disk. Claude Code's monitoring documentation describes the export as opt-in: it goes only to the collector you name. Each claude_code.api_request event carries a query_source attribute naming the subsystem that issued the request, a model attribute, token counts, and an estimated cost_usd value.

  1. Point the export at a collector you control. Set the variables from the monitoring documentation in the shell you start Claude Code from: CLAUDE_CODE_ENABLE_TELEMETRY=1, OTEL_LOGS_EXPORTER=otlp, an OTLP protocol, and an OTEL_EXPORTER_OTLP_ENDPOINT naming your collector. Do not turn on prompt, response, or tool-detail logging for this measurement: the cost fields are enough, and those extra switches would send conversation content to your collector.
  2. Work normally for a full working period. Run the mix of main-session work and background agents you actually run. A short test with one agent will not show the pattern. Include at least one long-running background agent with a large transcript if that is part of your normal work.
  3. Group the request events by query_source. In your telemetry backend, take the claude_code.api_request events for the period and group them by query_source, summing cost_usd and counting requests. Compare the agent_summary group with the main conversation group. Record three numbers: the summary share of estimated cost, the summary share of requests, and which model carried the summary spend.
  4. Read the result against your workload, not against either published figure. A small share with few background agents is an answer: stop there. A large share alongside many long-running background agents matches the Flamerix AI pattern. A large share with one short-lived agent would be a surprise worth raising with Anthropic, with the session details, before reaching for a workaround.

Price one call at the model's rates

Take the transcript size from your own export, or use a round figure to see the shape. At Anthropic's published API prices for Claude Opus 5.5, input costs $4.00 per million tokens, a cache read costs $0.20 per million tokens, and output costs $20.00 per million tokens. One summary call on a 200,000-token transcript therefore costs about $0.04 when the transcript is read from the cache (200,000 × $0.20 per million), and about $0.80 when it misses the cache and is billed as plain input (200,000 × $4.00 per million). The output is a few words, so at the measured ratio of roughly 9,000 input tokens per output token it adds well under a tenth of a cent. The point of the arithmetic is the gap between the two cases: one missed cache read costs about twenty times one cache hit, which is why a small miss rate can carry a large share of the bill. If your background agents run on a cheaper model, price the same call at that model's input and cache-read rates instead.

What to change, in order

  1. Update Claude Code first. Anthropic's reply says a fix for the most expensive case, summaries that miss the prompt cache, shipped in version 2.1.311 and later. Updating does not remove the recurring calls, but it addresses the case that carries the outsized cost in the Flamerix AI measurement. Measure again after updating, on the same kind of workload, before deciding anything else.
  2. Run fewer background agents at once. Each actively working agent gets its own summary call on its own timer, so the cost grows with the number of agents running in parallel and with the size of each transcript. If your measured share is large, two or three parallel agents is a reasonable place to start, then measure again. This is a decision about how you work, not a setting.
  3. Put background agents on the cheapest model that can do the job. The summary call goes to the agent's own model, so an agent running on the strongest model pays the strongest model's rates for every status line. An agent whose work does not need that model costs less in its own turns and in its summaries. Price both at the published rates before switching, and keep the strongest model for the agents whose work actually needs it.

What cannot be turned off

There is no supported setting that stops the summary calls while background agents keep running, according to Flamerix AI's write-up. Two environment variables do stop them, but only by stopping forked subagents or background agents altogether, which removes the work the summaries describe. Anthropic's reply says the team is working on a much cheaper mode that sends short notes to a smaller model instead of the full transcript. That mode was described as in progress on 29 October 2026, not as shipped, so it is not a step you can take yet.

Flamerix AI also publishes a third-party preload that intercepts the summary timer. It was tested on Linux across Claude Code versions 2.1.257 to 2.1.312, with a macOS report from another user, and it fails open: if a Claude Code update changes the code it matches, the summaries silently return. It changes how the client runs, it is not an Anthropic setting, and this page gives no install steps for it. Anyone considering it should read the write-up, including its warning that changing a client's runtime behavior may conflict with the terms of the service, and run its own checks again after every Claude Code update.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Token cost calculator · What an agent costs · Claude Code hooks token cost