Google AI Agent Completes Closed-Loop Wet-Lab Experiments

Google's official technical blog recently shared progress on its AI agent-driven autonomous wet laboratory. The team enabled the intelligent system to perform a closed-loop test on a liquid-handling robot: designing molecules, scheduling experiments, reading results, and adjusting the next round of plans. This moves molecular design from computer simulations to repeated validation in a real laboratory setting.

Wet labs are experimental facilities where reagents, biological samples, and liquids are handled through test tubes, pipettes, and other equipment, in contrast to dry labs that rely solely on computer simulations. Google notes that AI systems need to operate in wet labs to interact with the real world, since dry-lab results often deviate when applied to real environments.

The core approach is to let the AI system refine its judgments based on real biological feedback. The system not only runs experiments and reads outcomes, but also feeds results back into the model, allowing the agent to adjust its strategy in a loop. To help the agent understand the lab, the team built a framework called VERID. Based on an open-source robot control framework, VERID maintains a continuously updated digital twin of the workbench, tracking the robotic arm's position, liquid volumes in wells, reagent types, and pipette tip inventory. It checks for issues such as collisions, empty aspiration, or mismatched containers before sending instructions to the robot. Video and action logs from runs are also preserved for researchers to review.

Most existing automation equipment still relies on human-written programs and debugging. In this project, the team delegated the "review results and revise plans" step to the system, but the system must re-check equipment and sample states before making changes. In a myoglobin screening test, the system initially thought it had found a better molecule, but after adding controls it realized it had not truly identified the target. It then adjusted its design rules, and the next round improved.

Google emphasized that this is still a research prototype, not a mature product. Researchers still need to review sequences, verify plans, and supervise robot operations. The value of the AI agent lies mainly in compressing the tedious steps between prediction and validation, accelerating design and iteration in early-stage discovery.