New MIT paper: If agents can see and reuse what earlier agents built, the whole system gets stronger over time.

Even if an isolated agent still discovers the best individual solution.

The useful thing about a swarm is not necessarily better individual agents; it is letting their separate discoveries accumulate into a stronger shared system.

If your goal is cumulative improvement rather than 1 perfect answer, SwarmWorld suggests giving agents a persistent shared environment where useful work survives between agents.

So the paper is not saying that more agents are always better.

It is saying groups help when the task rewards accumulation: different agents discover different things, leave them behind, and let later agents build on that progress.