Google Research Puts Agents at the Wet-Lab Bench -- a Digital Twin Blocks Physically Impossible Robot Commands, and Agents Recover From a 'Sticky Binder' Trap in Three Rounds

On October 16, 2026, Google Research researchers Pooja Rao, Petar Sirkovic and Annalisa Pawlosky published "Grounding AI agents in biological reality", which summarises three papers on letting agents run real experiments. The motivation is stated plainly: "frontier models still do not reason reliably about 3-dimensional geometry and physical state", while wet-lab readouts are becoming "more of a bottleneck than ever" as in-silico design speeds up.

1. VERID: a digital twin between the agent and the robot

VERID (VERified Infrastructure for Discovery Laboratories) is "a semantic digital twin of the laboratory workcell", a live 3D tree that tracks robot arm geometry, well volumes, reagents and pipette tips. It is built on the open-source PyLabRobot framework. The post's core point is that "The safety of a command depends on the live physical context as well as the code": an agent can write valid code that lowers an 8-channel pipette into a trough too short for it, or aspirates from an empty well. VERID gives agents live state checks, fast virtual dry runs, hardware safety checks that "Block invalid commands before they reach the robot", and synchronized video and action logs.

In the case studies, the simulator rejected an 8-channel aspiration that would not fit. The agent replanned and designed a factorial diagnostic that isolated a subtle tip-wall carryover bias that replication alone "would have missed". In a self-play campaign reconstructing blinded dye mixtures from camera images, the agent improved its accuracy "more than fivefold".

2. Aptamers: the loop that caught its own mistake

Aptamers are DNA sequences that bind proteins, and a full design-test-analyse cycle takes 7 to 10 days. Over three rounds on two targets, agent-led design raised confirmed on-target hit rates to 33.5% for myoglobin and 42.1% for alpha-thrombin. The myoglobin run is the instructive one. In Round 2 the agents chased raw binding and filled 63% of designs with G-quadruplex motifs. The wet-lab data showed most were non-specific "sticky" binders that stuck to bare beads, and specificity fell from 75.0% to 19.1%. Reading that readout, the agents rewrote their design rules for Round 3, capping those motifs and subtracting bead binders. Specificity recovered to 72.4%.

3. Protein binders

For fibroblast activation protein alpha (FAP), a large receptor overexpressed in solid tumours, agent-guided "epitope excision" folds only the local binding pocket. It sped up each structure-prediction cycle 8.6 times. At equal compute, a scientist setting strategy with an agent handling tactics "nearly doubled the in-silico hit rate of a scientist working alone". The top designs bound with single-digit nanomolar affinity.

Safety design. Google says it chose non-infectious systems "without gain-of-function risk", that every designed sequence goes through biosecurity screening, that agents "never send unverified code directly to the liquid handling robots", and that humans authorise sequences and supervise robot runs.

What remains uncertain. These are Google's own reports, posted as PDFs on Google's servers; we found no peer-reviewed versions, and the targets were deliberately chosen to be tractable. The post lists open problems itself: making sure agents optimise "for true target binding rather than unintended experimental shortcuts", and verifying multi-step protocols.

Analysis: For anyone building agents that act in the physical world, VERID is the reusable idea: a stateful model of the environment that every action must pass through before execution, with the full trace logged. That is the lab version of a sandbox with egress rules. The sticky-binder episode is the most honest result in the post. The agent optimised the metric it was given, the physical readout exposed it, and the fix depended on the loop returning real data rather than a proxy score.