It has been discovered that Claude Code periodically sends transcripts from each background agent to its model.
Claude Code's 'Agent Summary' feature updates the background agent status approximately every 30 seconds, resending the entire transcript (recorded data) to the model. This transcript transmission consumes approximately 15-20% of the tokens.
Claude Code burns ~15-20% of your tokens on a 3-word status line: a Bun preload that turns agent_summary off · by Flamerix AI, flamerix.ai · MIT
Claude Code regularly sends each background agent's transcript to its model | Hacker News
https://news.ycombinator.com/item?id=50284255
Flamerix AI , a developer of AI tools, analyzed OpenTelemetry exports and found that the agent summary feature accounts for approximately 15-20% of Claude Code's total token expenditure, with roughly one in five API requests being sent for the agent summary feature. The cost increase was particularly pronounced when using the Opus model, with token consumption ranging from 18-26%.
Specifically, Claude Code sends the entire conversation log of an agent to its own model approximately every 30 seconds, requesting that it 'describe your most recent action in 3-5 words.' The responses received from each agent are used solely to update the status row in the agent panel and do not provide any information to the model or conversation. The agent summary feature has been enabled by default since Claude Code version 2.1.232 in August 2026 and cannot be disabled through settings.
To address the issue of not being able to disable agent summaries, Flamerix AI has proposed a solution by developing a Bun preload script called 'no-agent-summary.js'. 'no-agent-summary.js' is a JavaScript script of about 70 lines that wraps the global function setTimeout() and prevents the timer from starting when it detects a callback containing a specific anchor such as '[AgentSummary] Timer fired'. In other words, it disables only the timer related to the agent summary feature and does not affect other timer processing. It does not modify the Claude Code binary itself and is loaded using Bun's '--preload' option, so it will not be disabled by updates to Claude Code.
According to Flamerix AI, the impact and points to note regarding 'no-agent-summary.js' are as follows:
- Visible effect : Agent summary requests will be reduced to zero, and the panel status for each agent will no longer be updated.
- Overhead : Although Function.prototype.toString is called for each timer callback, the impact is minimal because it is cached in WeakMap (increase from approximately 10ms to approximately 17ms).
• OS compatibility : Operation has been confirmed on Linux, macOS, and Windows (however, Windows operation is limited).
・Note on updates : If the script anchor changes when Claude Code is updated, it may stop working, so you must perform the verification procedure after updating.
Furthermore, the Claude Code Lab , operated by the IT venture company Arrowheads , states that in addition to the aforementioned issues when the anchor is changed, it is loaded into all Buns on the machine in order to use the environment variable 'BUN_OPTIONS,' and therefore 'does not recommend' using 'no-agent-summary.js.'
Does Claude Code's sub-agent generate summaries every 30 seconds when running in the background? Measuring agent_summary token usage | Claude Code Research Institute
https://www.claude-code-lab.com/blog/claude-code-subagent-agent-summary
Furthermore, Opus has been reported to have another issue: it includes the user's email address in the User-Agent header of requests.
While we're modding Claude Code, Anthropic includes your email in the system pro... | Hacker News
https://news.ycombinator.com/item?id=50285046
Related Posts:
in AI, Posted by log1c_sh