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

Original Article ↗ Hacker News Discussion ↗

Claude Code’s built‑in status summaries for background agents are consuming a surprising share of users’ paid tokens, with some reports showing 8–16% of total usage going to periodic “agent_summary” calls that update a small status label. While Anthropic staff say telemetry shows a lower average impact and note recent fixes plus plans for cheaper summary modes, many users criticize the lack of a clear toggle, see a conflict of interest when model providers control both harness and billing, and experiment with hacks, alternative clients, or older binaries to cut waste. The episode feeds broader frustration with Claude Code’s growing bloat, opaque limits, and privacy quirks, and prompts comparisons with lighter third‑party harnesses and rival tools.

Background agent summaries & token burn

  • Core issue: Claude Code sends each active background agent’s full transcript to the model about every 30 seconds to generate short status summaries.
  • One heavy user measured ~16% of weekly token-equivalent spend from these calls; others report ranges from ~3–8% up to 15%+ depending on workflow and subagent usage.
  • Impact is worst for long-context agents (e.g., hundreds of thousands of tokens) and many concurrent subagents.

Scope, caching, and mechanics

  • Summaries apply only to actively working background agents, not idle ones. Regular single-session use is less affected.
  • Some testers say workflows (“ultracode”) are not impacted, only subagents.
  • Prompts are largely served from the provider’s prompt cache, so most are “cheap reads,” but repeated large cached reads still accumulate cost.
  • Cache TTLs: main sessions ~1 hour; subagents default ~5 minutes unless extended for a premium cache write. This makes cache behavior and cost somewhat complex and timing-dependent.
  • Users note aggressive auto-compaction near cache expiry reduces costs but can harm continuity or “feel like a lobotomy” for some workloads.

Provider response and planned fixes

  • A Claude Code engineer reports telemetry showing summaries at about:
    • ~2% of cost overall (API list prices),
    • just under 4% of requests,
    • ~6% of cost in the terminal, where most such calls occur.
  • They acknowledge it’s “a lot for a status line,” claim to have shipped a fix for cache misses (version 2.1.311+), and are working on a cheaper mode using short notes to a smaller model.
  • There is currently no official setting to fully disable summaries without also disabling background agents/forks.

User workarounds and tooling

  • A Bun preload script patches setTimeout to intercept and suppress the summary timer without modifying the binary; controlled tests show it drops summary calls to zero while keeping agents running. Windows runners and similar tools have been shared.
  • Some users freeze on older Claude Code CLI versions and disable auto-updates, claiming newer versions burn more tokens via background features.
  • Others rely on external orchestrators/“decision models” (e.g., Pi-based harnesses, omp.sh, custom bridges) to choose cheaper models and avoid Claude Code’s behavior.

Incentives, trust, and controversy

  • Many see this as unnecessary, non-transparent token burn that should be user-togglable or free, especially for enterprise/API-priced customers.
  • Opinions diverge on intent:
    • Some call it exploitative or even “fraud,” arguing model vendors owning harnesses have a built-in incentive to waste tokens.
    • Others think it’s poor engineering/product design: features built in an internal, cost-insensitive environment, then shipped without adequate cost controls.
    • A few argue competition and user churn limit how much deliberate waste would be rational.
  • Broader frustration surfaces about Claude Code’s perceived bloat, frequent behavior changes, opaque quotas, and a pattern of token-heavy features.

Comparisons to other harnesses and models

  • Several commenters prefer lighter or OSS harnesses (Pi, omp.sh, Codex, Muse Code, Copilot), claiming:
    • Better token efficiency,
    • More predictable behavior,
    • Less “spyware-like” or intrusive defaults.
  • Others say Opus + Claude Code still delivers exceptional productivity for them even at high token usage; experiences vary widely depending on project size, language, and workflows.

Privacy and safety concerns

  • Separate but related issue: Claude Code includes the user’s email in system prompts and sometimes in HTTP headers (User-Agent) when sites request contact info (e.g., SEC, Wikipedia, Crossref).
  • Some users patch this out via plugins; others argue the agent should always ask for consent before transmitting personal identifiers.
  • There is concern about prompt-injection scenarios where third-party sites could exfiltrate PII or trigger sensitive tool calls without explicit user approval; others see current leaks (e.g., email in UA) as minor but agree better mechanisms are needed.