Disable Claude Code Background Agent Summaries to Save Twenty Percent Tokens

01The Hidden Token Drain

By default since version 2.1.232, Claude Code runs an internal timer every 30 to 240 seconds for each active background agent. This timer gathers the agent's full transcript and instructs the model (typically Opus) to describe the latest action in three to five words for the agent panel's UI status line. In OpenTelemetry exports, these calls appear as claude_code.api_request events marked with query_source = agent_summary. In heavy multi-agent workflows, these status pings can account for 15% to 20% of your total token spend, directly eating away at your API rate limits.

02Why Standard Flags Fail

The undocumented environment flag CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES and the SDK setting agentProgressSummaries cannot turn the feature off: in the compiled binary, the parameter is part of an OR expression where internal paths force it to true. Furthermore, DISABLE_NON_ESSENTIAL_MODEL_CALLS does not exist in the binary, and modifying strings inside the executable fails because Claude Code runs as precompiled Bun bytecode.

03Suppressing the Timer with Bun Preload

The fix relies on BUN_OPTIONS=--preload pointing to a minimal wrapper (no-agent-summary.js). The script intercepts setTimeout and setInterval calls, inspects Function.prototype.toString() for the [AgentSummary] log anchor, and replaces the callback delay with an inert 2^31 - 1 millisecond wait. After saving the script, kill the daemon using claude daemon stop --any so that all newly spawned processes and pre-warmed background sessions inherit the hook.