Every 30 Seconds Claude Code Burns 18–26% on Agent Summaries

Claude Code automatically summarizes background agent activity every 30 seconds by sending each agent’s entire transcript to the agent’s model, requesting a 3-5 word description of the most recent action. This summary updates only the agent status panel and never reaches the user’s conversation. The feature has been enabled by default since version 2.1.232 in August 2026, with no UI setting to turn it off.

Independent measurements show agent summary requests consume 18–26% of Claude Code’s per-request API costs. Despite prompt caching reducing individual request costs, sending full transcripts repeatedly creates significant cumulative expense. A cheaper alternative introduced in version 2.1.310 uses 2k tokens of notes instead of full transcripts, but remains disabled by default and requires server-side configuration to activate.

The feature remains invisible to normal token-tracking tools because these API requests don’t appear in session logs—only OpenTelemetry exports capture them. Users can suppress the summaries using a Bun preload hook, though this requires manual setup and re-verification after each Claude Code update.