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

Who Should Read

A developer running background subagents in Claude Code whose subscription limits drain faster than expected

Core Mechanics

• It resends the full transcript every 30 seconds
  • For each background agent, Claude Code sends the entire transcript to the agent's own model (usually Opus) about every 30 seconds.
  • The only purpose is a 3-5 word status line in the agents panel.
  • The answer never goes back to the model or your conversation.
• No off switch, and on by default since 2.1.232
  • Since 2.1.232 (August 2026) it is on by default for interactive background agents.
  • The SDK option agentProgressSummaries has no effect when set to false.
  • The interval comes from a server flag, is clamped to 30-240 s, and the default is already the minimum.
• 16.4% of spend in the author's telemetry
  • query_source=agent_summary was 16.4% of all Claude Code spend, versus 15.7% for the main conversation.
  • It was 19.7% of all API requests, about 1 in 5.
  • 96% of that spend was on Opus, with about 9,000 input tokens per output token.
• A setTimeout wrapper kills only the summary timer
  • A Bun --preload script (about 70 lines) wraps the global setTimeout.
  • If a callback's source contains a summarizer anchor like "[AgentSummary] Timer fired", it returns an inert handle.
  • The binary is not modified, and all other timers pass through.
• With the hook: zero summary requests, frozen status line
  • On the author's machine (2.1.301) agent_summary requests dropped to 0 and no other side query grew.
  • Subagents and background agents run normally; only the panel status stops updating.
  • An update can change the anchors and silently re-enable summaries, so verify after every update.
• Prompt cache makes it cheaper, not cheap
  • About 98% of these calls re-read the transcript from cache.
  • The roughly 2% hitting a cold cache carry over a third of the cost.
  • These requests are not written to on-disk session logs, so only the OpenTelemetry export shows them.

Community Reactions

• Estimated 16% of spend from own transcripts
  • A 20x Max user had Opus analyze the gist and estimate from the 30-second timer.
  • They report $1,970 API-equivalent weekly spend and about $380 in hidden status calls, 16%.
• Reproduced at about 8% or less
  • One commenter says it affects only subagents, not workflows.
  • They say it is mostly cache reads, so short-context subagents cost little and long ones cost much more.
  • Another commenter measured about 3% across everything and everyone.
• Debate over whether cache hits make it cheap
  • Someone asked if cache hits make it cheap anyway.
  • The reply was that 16.4% comes from Claude Code's own cost_usd estimate, and 1 in 5 requests adds up.
  • Another commenter called the "15-20% of your tokens" claim unclear on its denominator.
• Opinion that model makers shouldn't own the harness
  • Several commenters argued inference providers have no incentive to optimize token use.
  • One says they decompile the binary to edit hardcoded system prompt instructions.
  • Another shared a Windows runner for measuring the traffic.

How to Apply

• Measure agent_summary share in OTel today
  • Enable OpenTelemetry export and group claude_code.api_request events by query_source.
  • Compare agent_summary request share and cost_usd against the main conversation.
  • If the share is small, skip the hook.
• Try the preload on a single session first
  • Download the file with the gist's curl command and run claude daemon stop --any.
  • Put --preload=<absolute path> in BUN_OPTIONS and launch claude.
  • Use the = form, because the space form breaks bun subcommands like bun install.
• Verify with CC_AGENT_SUMMARY_LOG
  • Set the log path and run a background subagent for a minute or two.
  • Check for the line 'summary timer #1 suppressed'.
  • A/B with CC_AGENT_SUMMARY_MODE=pass to confirm summaries return without the hook.
• Add a re-check after every update
  • After each Claude Code upgrade, re-check the log and confirm agent_summary stays at 0 in OTel.
  • Put BUN_OPTIONS in headless runs (-p), cron and CI too, and restart the daemon.

Try It

# Paste into a Linux terminal - install and verify commands from the original gist

curl -fsSL https://gist.githubusercontent.com/flamerix-ai/c3905705cc4205e5de3011fa2e74efce/raw/no-agent-summary.js -o ~/.local/share/no-agent-summary.js

claude daemon stop --any

# Check it loads: expect a 'preload active mode=off' line
BUN_OPTIONS="--preload=$HOME/.local/share/no-agent-summary.js" CC_AGENT_SUMMARY_LOG=/tmp/nas.log claude --version && cat /tmp/nas.log

# Run with the hook applied
BUN_OPTIONS="${BUN_OPTIONS:+$BUN_OPTIONS }--preload=$HOME/.local/share/no-agent-summary.js" claude

# Next: run a background subagent for 1-2 minutes and look for 'summary timer #1 suppressed' in /tmp/nas.log

Terminology

Related Papers

• Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)

  A KV pool smaller than the context window leaked other customers' data into 155 long-prompt responses.

• Microsoft Decision-1: 9B Qwen Model Launch Explained

  Microsoft's 9B Qwen-based model returns only per-option probabilities for routing and classification.

• Cerebras Deploys Trillion-Parameter AI Models

  Cerebras fits Kimi K3 (3T) across 59 wafers in a pipeline, serving 1000-2000+ tokens per second

• A practical guide to AI image generation, from diffusion to agents

  Testing image APIs showed you pay per canvas, and camera-style prompts beat describing the view

• DeepInfer: High-performance inference framework for DeepSeek models

  DeepSeek open-sourced DeepInfer, its production inference engine, under Apache-2.0.

• Mercury 3: Diffusion at scale

  Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality

• DeepSeek banned from U.S. federal civilian agencies under new DHS order

  DHS order bans DeepSeek from U.S. federal civilian agency IT systems

• Timestral: Mistral's time-series foundation models

  A 209M time-series model ranks first on three benchmarks and runs on CPU.

Related Resources

• https://gist.github.com/flamerix-ai/c3905705cc4205e5de3011fa2e74efce
• https://github.com/AlexErrant/claude-code-no-email
• https://news.ycombinator.com/item?id=50300097