Read a paper on multi-agent teams today, and the result surprised me a bit.
Three AIs solve math problems together. In one setup they have clear roles: one plans, one writes the answer, one reviews — nothing gets submitted unless the reviewer says yes. In the other, no roles at all: anyone can write, anyone can submit, and they find out right away if it's wrong, then try again.
To be precise: both sets of numbers below are from gemma-4 on Omni-MATH.
First try, both teams are about the same. The gap shows up after a wrong answer: the team with roles only recovered 22% of the time, the team without roles 61%.
The reviewer is the problem. Too cautious. A lot of answers never get submitted because they don't look good enough. No submission, no feedback, no idea what went wrong. Some questions were never submitted once.
Then the researchers changed one rule only: submit every round, keep everything else. Recovery jumped from about 20% to 61%.
So roles aren't the problem. Blocked feedback is. When checking is cheap and reliable, submit more. When it's expensive or unreliable, a strict gatekeeper makes sense.
Paper: Beyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent Systems, arXiv:2611.03931, submitted 2026-11-02.