On October 16, 2026, Google Research published “Grounding AI agents in biological reality: Building blocks for autonomous wet labs” by Pooja Rao, Petar Sirkovic and Annalisa Pawlosky, summarizing three projects with linked technical papers that connect AI agents to physical experiments.
VERID (VERified Infrastructure for Discovery Laboratories) is a digital twin of a liquid-handling robot workcell, built on the open-source PyLabRobot framework, that tracks arm geometry, well volumes, reagent identities and tip inventory, and provides virtual dry runs, hardware safety checks and synchronized video and action logs. In case studies an agent used the video and traces to notice uneven dye mixing and change its strategy, recovered from a mid-run plate-reader failure, and improved its accuracy more than fivefold in a self-play campaign reconstructing blinded dye mixtures.
In an aptamer sandbox, agents ran three design-test-iterate rounds of 7 to 10 days each on two protein targets. For alpha-thrombin, round 1 yielded no binders because fixed primer overhangs disrupted folding in 48% of sequences; hit rates then rose to 26.0% and 42.1%, at 98.5% specificity. For myoglobin, agents were initially misled by “sticky” G-quadruplex motifs (63% of round-2 designs), which cut specificity from 75.0% to 19.1%; in round 3 they revised their design rules and reached 72.4% specificity and a 33.5% on-target hit rate. A third project, epitope excision, sped structure prediction for binders against the tumor protein FAP by 8.6 times without losing accuracy; human-agent collaboration nearly doubled the in-silico hit rate versus a scientist alone at equal compute, and top designs reached single-digit nanomolar affinity.
Why it matters: most AI-for-biology results stop at computational predictions; this work closes the loop with physical robots, and the myoglobin episode shows agents catching and correcting their own systematic error from experimental data.
What it does not show: the aptamer work covers only three rounds on two targets, and the authors stress that humans still authorize sequences, verify protocols and supervise runs. Full methods and figures are in the linked papers rather than the blog post.