Designing the Sub-Agent Handoff So the Citation Agent Can Use It

Each of your five sub-agents currently hands back a structured summary of about 300 tokens: a finding statement, a confidence score, and a list of source URLs. In production the citation agent reports that it cannot verify roughly 40% of claims, because by the time it runs it only has URLs, several pages have changed or are paywalled, and re-fetching would mean new tool calls from inside Stage 3. An engineer proposes the obvious fix: hand back each sub-agent's full search-and-read transcript (about 30k tokens each) so nothing is lost.

1. Quantify why the full-transcript proposal fails, and name the failure mode from the subject's vocabulary that it reproduces.
2. Design the structured summary schema that serves both the supervisor's needs (aggregation, redundancy detection, outline) and the citation agent's needs, justifying each field against a specific downstream consumer in the pipeline.
3. Even with your schema, describe a handoff-loss failure in the usual direction (too little preserved) that would let a report pass faithfulness and citation precision while misrepresenting the evidence, and say which evaluation technique from Step 8 would catch it.