Claude Code regularly sends each background agent’s transcript to its model
Flamerix AI, which published the original analysis as a GitHub gist in late October, examined a feature in Claude Code, Anthropic's coding agent, called agent_summary. It fires for each background agent that is actively working, roughly every 30 seconds, and skips idle ones. Each firing sends the agent's full transcript to that agent's own model, usually Opus, with a request to describe the latest action in 3-5 words. The reply fills a status label in the agents panel and, according to the gist, never reaches the model or the user's conversation.
In Flamerix's own week of measurement, these calls made up 16.4% of spend, slightly more than the 15.7% attributed to the main conversation. The gist uses Claude Code's per-request cost estimate rather than an invoice, so on a subscription the cost appears as lost rate-limit headroom. Sesame Disk reports that 96% of the summary spend ran on Opus. The behavior has been on by default for interactive background agents since version 2.1.232, which arrived in August 2026.
There is no official off switch. Flamerix ships a Bun preload script, no-agent-summary.js, under the MIT license, that replaces the summary timer with an inert one. The gist warns that the script can fail silently after a Claude Code update. Environment variables that appear to help either do nothing or also shut off subagents and background agents.
Why it matters
For teams running several background agents, cost now scales with how many agents are active and how long their transcripts grow, not only with the prompts they receive. A long-running agent with a large context gets resent in full every 30 seconds while it works. That makes per-agent unit economics harder to forecast, and it means a coding product's gross margin can depend on overhead that the model vendor sets and can change in a point release.
For investors and buyers, the episode shows that vendor telemetry and local logs can disagree. The summary calls do not show up in session logs on disk, so teams relying on local files will undercount spend. Buyers evaluating agent tools should ask vendors for per-request cost breakdowns. Founders building on a model provider should budget for background calls they do not trigger directly, and measure them before scaling agent fleets.
Key points
- 1Status calls run only for background agents that are actively working, roughly every 30 seconds.
- 2Each call sends the agent's full transcript to that agent's own model, usually Opus.
- 3Flamerix measured the feature at 16.4% of one week's spend, versus 15.7% for the main conversation.
- 4The behavior has been on by default for interactive background agents since version 2.1.232.
- 5
