Claude Code regularly sends each background agent’s transcript to its model
Investigators and a self-identified Claude Code team member described recurring model calls for active background-agent summaries, with costs varying by workload.
TL;DR
- Flamerix AI reported that Claude Code resends active background-agent transcripts to generate short progress summaries.
- The team attributed 16.4% of estimated spending in one measured week to the calls, while a self-identified Claude Code team member reported about 2% across telemetry-sharing users.
- Independent tests covered selected versions, and users disputed whether a proposed setting disables the calls.
Flamerix AI reported that Claude Code sends an active background agent’s transcript back to its model to generate a short progress label. A Hacker News user who said they work on Claude Code confirmed the basic mechanism and said idle agents do not receive the calls. [1] [2]
Flamerix AI attributed 16.4% of estimated Claude Code spending in one measured week to these summaries. The self-identified team member said the share was about 2% at API list prices across users who send telemetry, and about 6% in terminal use. [1] [2]
The AI News Report separately inspected the macOS binary for Claude Code 2.1.309 and found the summary prompt and timer-related strings, without running an API test. In a GitHub issue, a user reported five summary calls roughly 32 seconds apart in a Linux test of version 2.1.312. [3] [4]
The measurements cover particular workloads and versions, so they do not establish every user’s cost. In the GitHub issue, users reported conflicting results for a proposed setting, leaving its effect unresolved. [2] [4]
Why it matters
Agent summaries add model requests and resend developer-session context. Visibility into that background activity matters for teams assessing the cost and data flows of AI coding tools.
Editor's note
Flamerix AI’s percentage is an estimated cost share for one workload, not a token share or universal bill. The team member’s figure covers users who send telemetry; settings results are disputed.