Daily harness signal

Put a meter before the swarm

Claude Code now exposes hard session ceilings for searches and subagent launches. The shipped 200-call defaults are circuit breakers, not sensible budgets for most unattended jobs.

July 18, 2026 · JST Fresh · source date July 17 One implementation finding
Implementation lesson: the shipped default is only an emergency ceiling. Set workload-specific search and delegation budgets before launch, then prove the next call is refused at the tool boundary.
01 · Fresh · source date 2026-07-17 UTC

Replace open-ended autonomy with a tested call budget

Use when. A paid, unattended, or recursive Claude Code session can invoke WebSearch or spawn subagents—especially research, repository audits, verification panels, and resumed background work.

Action. Upgrade to Claude Code 2.1.212 or later. Derive each ceiling from the planned fan-out plus small retry headroom; do not inherit 200 by accident. For an unattended research job expecting six workers and about ten searches, launch with:

CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION=12 CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION=8 claude

Put the variables in the parent process environment for cron or CI; prompt prose is not enforcement. Before rollout, run two disposable low-cost canaries: set one ceiling to two, keep the other above the test, and explicitly request three sequential calls.

Acceptance check. In each canary, the transcript records no more than two WebSearch calls or two new subagent spawns; the third request returns a limit signal without creating another tool event or subagent transcript. Any third dispatch, silent retry, or continued fan-out blocks deployment.

Evidence. Anthropic’s official 2.1.212 release and Week 29 documentation name both variables, their 200-call defaults, and the runaway-loop purpose. Independent versioned issue artifacts show the failure shape: issue #68110 records one research request producing 48-plus agents and 1.5-million-plus tokens; #77060 preserves a July 13 resume-path reproduction with nearly 100 agents and 256-plus task files.

Caveat. These are lifetime call-count caps, not concurrency, nesting-depth, per-agent-turn, token, wall-clock, or monetary budgets. They do not cover WebFetch, shell networking, or MCP. Anthropic explicitly says /clear resets the subagent budget, so treat a clear as a new allowance and retain an external spend ceiling. The closed-source release publishes no regression-test output; keep the canary as a deployment gate.

Compact source notes

  1. Claude Code v2.1.212 (published July 17, 2026 00:26 UTC / 09:26 JST). Official shipped configuration and reset semantics.
  2. Claude Code Week 29 notes (July 13–17, 2026). Official summary independently names both caps and defaults.
  3. Claude Code issue #68110 (opened June 13, updated July 1, 2026). Versioned fan-out reproduction, transcript-derived scale, and counter-analysis.
  4. Claude Code issue #77060 (opened July 13, 2026). Fresh resume-triggered recursion reproduction with task JSON and turn metrics.