Ruleloaded every turn
Research and verification
- Web search is capped per session — 200 calls on Claude Code — and shared by every subagent. Budget roughly 10 agents at 20 searches, or 6 at 30, and stagger bigger runs across sessions. After a fan-out, check each subagent’s count and re-run starved dimensions in a fresh session.
- Prove a filter bites: send a value that can never match and confirm the result is empty.
- Spot-check one returned record against the field you filtered on before trusting the set.
- A swallowed per-item error is unknown, not absent. Count failures separately from empties.
- Re-check the live spec before assuming a documented workaround is still needed.
- Prefer primary sources. Trace every rule you rely on to a quotable source; when a file’s
header and body disagree, find the code or the spec that decides it. See
api-verification.