Claude Code regularly sends each background agent’s transcript to its model

569 points · 232 comments on HN · read original →

Points and comments are a snapshot, not live.

Claude Code's agent summary feature consumes 15-20% of token spend without a user-facing setting to disable it.

A Bun preload script suppresses the agent_summary timer that sends the full transcript to Opus every ~30 seconds for a 3-5 word status line. Before suppression, agent_summary accounted for 16.4% of Claude Code's estimated spend and 19.7% of API requests, with a 9,000:1 input-to-output token ratio. The hook wraps setTimeout/setInterval to block only the summary timer, preserving background agents and forks. The default is on since version 2.1.232 (August 2026). The documented switch (agentProgressSummaries) only enables it further. The script requires environment setup and verification after each update.

A cheaper 'short' mode exists in version 2.1.310 but remains off by default and requires server flags. The prompt cache mitigates but does not eliminate the cost.

What commenters are saying

Multiple commenters confirm similar token usage increases and frustration with the lack of a disable option. Some note that the impact depends on session patterns: single-session users see less overhead those running multiple background agents. A subthread debates whether prompt caching makes the cost negligible, with counterpoints citing the 5-minute cache TTL for subagents and repeated cache reads adding up. One user reports their 5-hour limit dropped from 25% to 13% weekly usage recently, attributed to aggressive compaction enabled by default. Another claims they must run on the Max plan to sustain heavy usage and have seen no relief. A separate discussion contrasts Claude Code's approach with lightweight alternatives like Pi.