AutoBioLab: Multi-Agent LLM System for Automated Biology Labs
An LLM agent framework that takes a scientist's intent all the way to executed experiments on liquid-handling robots, with checks before a run and automatic recovery when something goes wrong.
Motivation
Lab automation hardware is capable, but turning a scientific goal into a correct, machine-executable protocol still takes expert effort, and runs fail in ways that are hard to anticipate. AutoBioLab asks how far LLM agents can close that gap safely.
Approach
The system splits the work across specialized agents that plan the experiment, verify the generated protocol, and monitor execution on the liquid handlers. Validation checks catch command errors before anything runs; during execution, a hybrid recovery agent combines hand-written rules with LLM tool calls to detect and repair failures.
Evaluation
- A 48-task planning suite built from real experiments.
- A 96-scenario fault-injection simulator for testing recovery.
- Three rounds of testing, compared against 8 LLMs and agent baselines.
The paper is currently under review.
Related publications
- Under reviewAutoBioLab: From Scientific Intent to Machine Executable Laboratory AutomationUnder review, 2026* Equal contribution