Claude Code wastes 15–20% of tokens on status labels
Original: Claude Code burns ~15–20% of your tokens on a 3-word status line
Why This Matters
Token waste at scale hits rate-limit headroom for every Claude Code user running agents.
A Flamerix AI GitHub Gist reveals Claude Code sends full agent transcripts to Opus every ~30 seconds just to generate 3–5-word status labels, consuming 15–20% of all token spend with no user-facing benefit.
Claude Code's agent_summary feature, enabled by default since version 2.1.232 (August 2026), polls every active background agent every ~30 seconds. It sends the agent's entire conversation transcript to its model—usually Opus—and asks it to describe its most recent action in 3–5 words. That answer only updates a status line in the agents panel and never reaches the main conversation or the model context.
Flamerix AI measured the impact using Claude Code's OpenTelemetry export. Before applying a fix, agent_summary accounted for 16.4% of total Claude Code spend and 19.7% of all API requests—roughly 1 in 5 calls. The main conversation itself was only 15.7% of spend. About 96% of agent_summary spend ran on Opus, with an input-to-output token ratio of approximately 9,000:1. An independent GitHub issue (#86879) collects four separate measurements ranging from 18–26%.
Flamerix released a ~80-line Bun preload script (MIT license) that wraps setTimeout and setInterval. When it detects a callback tied to the summarizer anchor, it returns an inert handle instead of arming the timer. All other timers pass through unmodified. The binary itself is not touched; BUN_OPTIONS=--preload injects the hook. The gist notes that a Claude Code update can silently re-enable the behavior, so users need to verify after upgrades.