According to Beating monitoring, a joint team consisting of Shanghua Gao, Ada Fang, and Marinka Zitnik from institutions such as Harvard Medical School, Kempner Institute, and Broad Institute has open-sourced the scientific discovery agent system AutoScientists. The system is based on the AI agent social collaboration platform ClawInstitute and operates without a central planner or orchestrator, simulating decentralized collaboration in real human scientific research. Previous systems like Autoresearch employed single-threaded hill-climbing searches, retaining changes only when immediate improvements were evident, which often led to bottlenecks after a few iterations. In contrast, AutoScientists introduces a decentralized forum mechanism where multiple sub-agents, encapsulated by Claude Code, exchange peer review comments through postings before consuming computational resources, thus avoiding redundant testing of failed paths. Multiple agents can also spontaneously form research groups around promising directions for multi-faceted exploration, addressing the blind spots of solo searches. In the training optimization of the language model GPT nanochat, the system achieved a 1.9 times increase in experimental speed; even starting from a strong baseline where further optimizations were not possible, AutoScientists completed 7 improvements, while the previous baseline had zero improvements. In the BioML-Bench test, which covers biomedical large model benchmarks in medical imaging, drug development, protein engineering, and single-cell genomics, the system achieved an average leaderboard percentile of 74.4% across 24 tasks, an increase of 8.3 percentage points over the previous strongest agent record. In the ACE2-Spike protein binding prediction, the Kermut extension method discovered by the system improved the Spearman correlation coefficient by 12.5%. When the same method was applied without modifications to all 217 ProteinGym evaluations, the official average Spearman correlation coefficient increased from 0.657 to 0.700, a rise of 6.5%, surpassing the overall best performance of the previous ProteinGym supervised benchmark.