Google AI agents run closed-loop lab experiments, in research

In a wet lab, a liquid-handling robot moves reagents from tube to plate, tracks how much liquid is left in each well and runs molecular biology protocols step by step. Google Research says today's frontier AI models still do not reason reliably about three-dimensional geometry and physical state, and that gap keeps an agent from running that robot on its own. Its blog post describes three projects aimed at closing it.

The first, VERID (VERified Infrastructure for Discovery Laboratories), is a digital twin of the lab workcell: a live, stateful 3D model that tracks the robotic arm, the volume in each well, the reagent identities and the pipette-tip inventory. Google Research notes that a command can be valid code and still be physically impossible, such as lowering an 8-channel pipette, a head with eight tips working at once, into a trough too short for it. An agent can run a virtual dry run in seconds to catch collisions or empty wells before touching reagents, and hardware checks block invalid commands before they reach the robot. The framework is built on PyLabRobot, an open-source framework, and records synchronized video and action logs for every run.

The second project tests the whole loop. Aptamers are short DNA strands that bind a target protein, and a full design-test-analyze cycle takes just 7–10 days, because DNA can be chemically synthesized in pooled libraries and read out directly by sequencing. Agents designed candidate sequences, screened them against target proteins on magnetic beads to isolate those that bind, and sequenced the results to feed the next round. Across three rounds on two targets, myoglobin and α-thrombin, confirmed on-target hit rates rose to 33.5% and 42.1%, respectively.

Each target hid a different wet-lab failure. On α-thrombin, Round 1 produced 0% binders, because fixed primer overhangs, the extra sequence tails added to each strand, disrupted folding in 48% of sequences. Clamped flanking stems in Round 2, then more diversified designs in Round 3, lifted the hit rate to 26.0%, then 42.1% at 98.5% specificity. On myoglobin, agents that prioritized raw binding affinity in Round 2 pushed G-quadruplex motifs, a DNA folding pattern, to 63% of designs. Raw pull-down rose, but the lab data showed most of those binders were not specific to the target.

These are Google Research's own results, from two protein targets over three rounds, and the post reports no independent replication. The team states that computational models cannot replace empirical validation, which is why agents need real experimental readouts. VERID's case studies are demonstrations, not a product. For a research group running biomarker or drug screens, the promise is faster iteration: a cycle of 7–10 days, with a robot that checks its own moves before running them. The post gives no date, price or access plan for either system, and it does not say when agents will run such experiments outside Google's research. Its stated aim is for agents to eventually separate biological signal from experimental noise across many candidates in parallel.