LabAgent is a reproduce‑and‑discover harness built for a lab’s ongoing work. It guarantees that every experimental skill can be run and verified, and it logs corrective actions and experiences for direct fixing or avoidance of similar mistakes. The first mechanism is a skill wrapper that encapsulates prerequisites, inputs, execution scripts and validation steps in a unified description. The second mechanism is an error‑log system that automatically captures exceptions and provides replayable correction commands. We deployed LabAgent on drug property prediction, biomedical problem analysis, protein variant effect prediction and statistical genetics. In each domain LabAgent outperformed commercial generalist agents and accurately reproduced a figure from a published study. These results show that LabAgent can effectively integrate existing laboratory knowledge and expand it in a reasonable manner.
Review