After the web-search agent finds 25 sources containing 120,000 tokens of raw content, the document-analysis...

Option B gives the report generator the two forms of information it actually needs: a compact narrative containing the synthesized conclusions and precise evidence records supporting those conclusions. The source index should map claim identifiers to URLs, document locations, excerpts, and other provenance fields so citations can be generated and verified without reloading 120,000 tokens of raw material.

Anthropic’s context-engineering guidance recommends distilling extensive subagent activity into high-signal summaries while preserving critical context and dependencies. Its research architecture similarly treats subagents as intelligent filters that compress large evidence collections before returning relevant findings to the coordinator.

Option A lacks sufficient location-level evidence for reliable citation verification. Option C postpones attribution until after drafting, when claim wording may no longer map cleanly to the original evidence. Option D maximizes theoretical completeness but introduces extensive irrelevant material, higher token costs, slower processing, and greater context pollution. The synthesis draft plus a structured claim-to-source index maintains evidentiary completeness where it matters while eliminating intermediate searches, duplicate passages, and other low-value context.

================