Agent teams learn reusable ways to organise their reasoning

What's this about?

Self-Organizing Agent Teams learn patterns for roles, turn-taking, challenge and synthesis from earlier collaborations, then reuse that organisation on unseen problems.

What exactly is demonstrability?

It measures whether a group can recognise a correct solution once someone produces it. Team discussion helps most when good reasoning can be identified and adopted rather than merely generated.

Why does this matter?

1. Teams average 66.7% versus 48.8% for their strongest member across five benchmarks.
2. They also beat compute-matched solo inference and a perfect router over independent answers.
3. Improvement tracks demonstrability strongly (Spearman rho 0.90), suggesting collaboration is task-dependent rather than universally helpful.