Is the 15–20% figure credible?
It comes from Flamerix's own setup and several users measuring heavy Subagent sessions, with numbers between 15% and 26%, while one person measured about 8% at most in their sessions.
What is credible is that overhead is significant under heavy subagent use, not that everyone loses 15–20%, and there is no Anthropic data.
I only run a single main agent. Am I affected?
agent_summary is a request issued for background agents, so an ordinary session with one main agent should in theory see a much smaller share.
I found no public measurement for that case, so to be sure, measure your own numbers with OpenTelemetry.
Is it safe to turn it off with the Bun preload script?
I would not use it in production as is. It was tested only on Linux with versions 2.1.257 to 2.1.312, macOS and Windows are unverified, and it runs with your privileges.
When an update changes the source it fails silently, so you may think it is off when it is not, which means remeasuring after every upgrade.
Is there an official way to turn it off?
I found no official setting or environment variable, and no Anthropic response. Someone asked whether the disableAgentView setting stops these requests, but nobody reported a measurement.
If you want an official switch, you can follow issue #86879.
On October 28, Flamerix AI published a GitHub gist claiming Claude Code spends about 15–20% of its tokens generating three-to-five-word status labels for background agents, and the claim spread quickly. This piece separates what is supported, what holds only in some cases, and what has no evidence, and gives you a way to measure your own number.
As the gist describes it, agent_summary is a side request that runs on a timer: for each active background agent, it sends that agent's transcript to the agent's own model and asks for a three-to-five-word description of the latest action, with the result used only for the agents panel status display, not the main conversation. The gist says it has been on by default since version 2.1.232 (August 2026). As early as August 15, a user had filed GitHub issue #86879 asking for a supported way to disable the agent_summary spinner.
Flamerix, over one week on its own setup: 16.4% of estimated spend and 19.7% of requests, falling to 0% after installing its hook. In the issue, one 4-hour session with 27 subagents and 1,496 requests: 22.2% of raw tokens, 18.0% weighted by Opus list prices, and 22.7% of requests. Other commenters reported weighted shares of 15–22% in heavy Subagent sessions, and the gist cites four measurements of 18–26%. On the other side, a Hacker News tester said it was about 8% at most in their sessions. These are all users' own measurements; there is no Anthropic data.
Supported from several directions: background agents do issue this kind of periodic summary request. Partly supported: under heavy use with many subagents, the overhead is significant. Not supported: treating 15–20% as the share of everyone's tokens; an ordinary session with a single main agent should see far less. I found no Anthropic response, and nobody has verified whether the company will add an official off switch.
No official setting or environment variable has been found. The gist offers an unofficial approach: load a script with BUN_OPTIONS=--preload that wraps timers and suppresses the summarizer's timer when it matches a source-text anchor. It was tested only on Linux with versions 2.1.257–2.1.312, fails silently when an update changes the code, and runs with your privileges. One commenter said an environment variable worked on 2.1.241 but not 2.1.245, and another mentioned an undocumented name of unknown behavior in 2.1.300; neither is confirmed. Someone asked whether the disableAgentView setting stops these requests, and nobody reported a measurement.
Export OpenTelemetry events, group claude_code.api_request by query_source, and compare tokens and requests for agent_summary against the main conversation. That is more reliable than trusting any single percentage.
First decide whether you are the affected kind of user: the overhead is likely to matter only for teams that routinely run many background subagents, while individuals with single sessions can mostly ignore it. Run the measurement above for a week; if the share is beyond what you accept, then decide whether to risk the unofficial script or reduce the number of concurrent background agents. Until Anthropic provides an official switch, do not put a preload script tested only on Linux into a production environment.