Parallel web-research fan-out: partial JSON per subagent, merge + normalize + spot-check

Large web-verified research tasks (a dozen entities, each needing several dated episodes plus a current-state snapshot, every number sourced) blow the context of a single agent and take too long serially; naive delegation returns prose that is hard to merge and unverifiable.

Write one shared brief file (schema, rules, per-entity hints, call budget, output path) into the scratchpad; spawn 3-4 general-purpose agents in one message, each owning 3 entities and writing an unwrapped JSON object to its own scratchpad file, then replying with only the entity list plus what could not be sourced. Orchestrator meanwhile gathers group-level context nobody else covers. On return: print a per-entity summary (episode counts, source counts, key numbers), normalize schema drift (e.g. "ongoing" -> null with a flag; drop entries agents self-flagged as failing a rule), independently re-fetch 2-3 headline numbers as spot-checks, record intraday drift in notes, then merge with a Python script that asserts every entry has a source and every required field is present before writing.

Splitting by entity keeps each agent under its call budget, a fixed schema makes the merge mechanical, and the orchestrator's own spot-checks are what make the compiled numbers trustworthy.

Any multi-entity research where each number must carry a URL and the caller will audit the output.