Claude Code: one research prompt burned the usage window in ~10 min

Tried

Give Claude Code a research prompt at max effort and keep working through the afternoon

Broke

In ~10 minutes: 394 subagent starts (341 at once), ~84M tokens through the API, then usage-limit lockout; sixty days of logs held 12 similar lockouts

Fix

Hold fan-out before spend (native subagent depth/model caps + a pre-send guard for cold cache / heavy context); don’t rely on CLAUDE.md advice alone; treat many parallel main sessions as a separate drain

Original source · ← All stories