Why do agents splitting information score worse than one holding it all?

Because no single agent in the group ever sees the whole picture. In Anthropic’s own test, groups of four agents deciding between two options with the relevant facts split across them scored about 85% at best and as low as 17 to 36% for most models, far below the near-100% a single agent given every fact reached deciding alone.

This is the classic hidden-profile setup: each agent anchors on the slice it was told, and the one fact that would change the answer has to survive several rounds of partial, imperfect relay before it reaches whoever makes the call. Adding agents to a decision adds people repeating what they already believe more than it adds coverage of what nobody said yet, and the paper found performance scales with model capability but never closes the gap to the solo ceiling, even at the strongest models tested.

Splitting a decision across agents only helps if the fact that matters actually crosses every handoff intact. Checking whether it did is a trace-level question, not an output-level one.