AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation

Researchers introduced AutoScientists, a decentralized multi-agent system for automated scientific research.

It shares experimental state, forms teams around promising hypotheses, critiques proposals, and learns from failed paths.

AutoScientists outperformed prior agents on BioML-Bench, GPT training optimization, and ProteinGym.

The results suggest collaborative, self-organizing agent teams can beat single-agent and centrally planned approaches in long-running research workflows.