A team from Harvard Medical School, the Kempner Institute, and the Broad Institute — led by Shanghua Gao, Ada Fang, and Marinka Zitnik — has open-sourced AutoScientists, a scientific discovery agent system. Built on the AI agent social collaboration platform ClawInstitute, AutoScientists has no central planner or central orchestrator. Instead, it mimics the decentralized collaboration of real human research.
Previous systems like Autoresearch used single-threaded hill-climbing search, only keeping changes that showed immediate improvement. That approach tends to stall after a few rounds. AutoScientists introduces a decentralized forum mechanism: multiple sub-agents, each wrapped by Claude Code, post peer reviews for one another before actually burning compute cycles, avoiding repeated testing of dead-end paths. Agents can also spontaneously form research groups around promising directions, exploring multiple paths instead of blindly forging ahead alone. When optimizing the GPT nanochat language model, the system achieved a 1.9x speedup in experiments. Even when starting from a strong baseline that a single agent couldn't improve further, AutoScientists still made 7 improvements — compared to zero for the previous baseline.
On BioML-Bench, a biomedical AI benchmark covering medical imaging, drug discovery, protein engineering, and single-cell omics, AutoScientists scored a 74.4% average leaderboard percentile across 24 tasks, beating the previous best agent by 8.3 percentage points. On ACE2-Spike protein binding prediction, a method the system discovered — called Kermut extension — improved the Spearman correlation by 12.5%. The same method, transferred without modification to all 217 ProteinGym evaluations, raised the official average Spearman correlation from 0.657 to 0.700, a 6.5% increase that surpassed the previous overall best for the ProteinGym supervised benchmark.