Claude Code Agent Summaries Burn 16–20% Tokens
Flamerix AI released an MIT-licensed Bun preload that disables Claude Code’s background agent summaries, which resend full transcripts every 30 seconds to generate cosmetic three-to-five-word status updates. Its telemetry measured agent_summary at 16.4% of spend, while related reports cite 18–26%.
This exposes a costly design flaw in agent UX: a tiny status indicator can quietly consume nearly as much budget as useful work.
- –Each summary call re-sends the agent’s full context, typically to Opus, despite producing only cosmetic status text.
- –The requests use skipTranscript, so transcript-based cost tools undercount the real usage.
- –The hook suppresses matching timers through BUN_OPTIONS while leaving normal agent execution intact.
- –It is a brittle, unofficial workaround: source-text matching can fail after updates, and Linux is the only tested platform.
- –Disabling summaries removes live agent-panel updates, making Anthropic’s missing supported opt-out especially notable.
DISCOVERED
1h ago
2026-10-29
PUBLISHED
2h ago
2026-10-29
RELEVANCE
AUTHOR
NikZaal