Google Research says it has begun connecting AI agents to physical lab robots so they can design, run and learn from experiments in tight loops. The October 16, 2026 post, Grounding AI agents in biological reality: Building blocks for autonomous wet labs, by Pooja Rao, Petar Sirkovic and Annalisa Pawlosky, presents three projects and headline numbers: aptamer hit rates up to 42.1% after three cycles, and protein binders with single-digit nanomolar affinity for a cancer target.
This is the "autonomous wet lab" idea moving from slogan to engineering. The interesting part is not a single discovery but the safety and verification plumbing that lets a language-model agent touch hardware. This post explains each piece, what the numbers do and do not show, and how it compares with other labs' agentic-science work we have covered.
TL;DR: the three building blocks
| Project | What it is | Headline result (Google's numbers) |
|---|---|---|
| VERID | Live 3D digital twin of a lab workcell on PyLabRobot | Dry runs in seconds; blocks invalid robot commands; accuracy on blinded dye mixtures improved more than fivefold in a self-play campaign |
| Aptamer sandbox | Closed design-test-sequence loop, 7-10 days per cycle | alpha-thrombin hit rate 0% to 26.0% to 42.1% at 98.5% specificity; myoglobin 33.5% on-target in round 3 |
| Epitope excision | Fold only the binding pocket of a large target | About 8.6x faster structure prediction; single-digit nanomolar FAP binders |
Why frontier models need grounding
Google frames two challenges. First, frontier models do not reliably reason about 3D geometry and physical state, which is exactly what a liquid-handling robot needs. Second, real wet-lab experiments have not kept pace with in-silico design: a model can propose thousands of molecules, but the bench decides which are real.
The stated long-term goal is for agents to "ideate and analyze, but also run lab experiments to test their hypotheses," learning from results while humans keep scientific agency. That framing matters. It is a supervised-loop story, not a claim that a lab runs itself.
VERID: a digital twin that says no to the robot
VERID is described as a live, stateful 3D volumetric entity tree. It tracks the geometry of the robotic arm, individual well volumes, reagent identities and the pipette tip inventory. Google lists four capabilities.
- Live 3D state tracking. The agent checks deck layout, well volumes and tips before acting.
- Fast virtual dry runs. Protocols are simulated in seconds to catch collisions or empty wells.
- Hardware safety checks. Invalid commands are blocked before they reach the robot.
- Synchronized video and action logs. Complete run records support reproducibility and debugging.
Four case studies from the post
- Video and execution traces revealed uneven dye stratification caused by dispensing too close to the well floor, and the agent refined its mixing strategy.
- The digital twin let an agent recover from a plate-reader failure mid-run without re-pipetting.
- The simulator rejected an 8-channel aspiration that would not fit a narrow trough. The agent replanned and designed a factorial diagnostic that exposed a bias from tip-wall carryover, which simple replication would have missed.
- In a self-play campaign, an agent reconstructed blinded dye mixtures from overhead camera images, identified its own optical and fluidic artifacts, and improved accuracy more than fivefold.
For software engineers the analogy is clear: VERID plays the role of a type checker plus a staging environment. It is the same pattern as sandboxing a coding agent, applied to pipettes. The paper is linked from the post at gstatic.com, and the base framework is documented at PyLabRobot.
The aptamer sandbox: what three cycles looked like
Aptamers are short DNA strands that bind proteins. Google used them as a fast prototype for closed-loop agentic discovery because DNA is chemically synthesized in pooled libraries and read out by sequencing. Each round, agents design candidates, screen them against target proteins on magnetic beads, and sequence the binders to feed the next round. Each cycle takes 7 to 10 days, and the team ran three. The paper is here.
alpha-thrombin, the harder target
- Round 1: 0% binders. Fixed primer overhangs disrupted folding in 48% of sequences.
- Round 2: the agents introduced clamped flanking stems and the hit rate rose to 26.0%.
- Round 3: diversified designs reached a 42.1% hit rate at 98.5% specificity.
Myoglobin, and a cautionary dip
In round 2 the agents prioritized raw affinity and enriched G-quadruplex motifs, which made up 63% of designs. Most turned out to be non-specific "sticky" binders. Specificity fell from 75.0% to 19.1%, and the bead-corrected hit rate fell from 10.2% to 3.7%.
In round 3 the agents capped promiscuous motifs, subtracted bead binders and moved to non-G4 scaffolds. Specificity recovered to 72.4% and the on-target hit rate reached 33.5%, which the post calls a ninefold increase.
The dip is the most instructive part. It is a textbook case of an optimizer exploiting a measurement shortcut, and the loop caught it. Google lists this as an open problem: making sure agents optimize for true target binding rather than experimental shortcuts.
Epitope excision and the FAP binders
The third project targets fibroblast activation protein alpha (FAP), a large receptor overexpressed in solid tumors. Designing protein binders adds wet-lab steps compared with aptamers because proteins need cellular or cell-free expression.
Re-folding dominates GPU compute and scales quadratically with target length. Excising only the local binding pocket during folding sped up structure prediction about 8.6x, and the text says "more than eightfold," without losing structural accuracy. It also turned pocket-selection into a set of tactical levers an LLM agent can tune.
On human-agent collaboration, Google reports that at equal compute a scientist setting strategy with an agent executing tactics nearly doubled the in-silico hit rate compared with the scientist working alone, and found hits across a broader range of target sites. Yeast surface display with fluorescence-activated cell sorting then screened large libraries in parallel, and the top designs yielded binders with single-digit nanomolar affinity for FAP. The paper is here.
What is claimed and what is not shown
- All numbers are Google's own, from the three linked papers and the blog; the post does not report independent replication.
- The aptamer targets are well-studied proteins chosen as a sandbox. The post does not claim a new drug candidate.
- "Single-digit nanomolar" is a binding affinity for FAP, not evidence of therapeutic effect.
- The post does not link standalone code for VERID or the other two projects, so outside labs cannot yet reproduce the stack end to end, though the PyLabRobot base is open.
- Comparisons to human-only baselines exist for the in-silico hit rate; the post does not give a human-only wet-lab baseline for the aptamer cycles.
Safety design
Google ties the work to its Frontier Safety Framework and bioresilience approach, and lists three measures.
- Safe testbeds. Well-controlled, non-infectious systems with no gain-of-function risk, and all designed sequences go through biosecurity screening.
- Simulation before execution. Agents never send unverified code to robots; every protocol is dry-run against the digital twin.
- Human oversight. Humans authorize and review sequences, verify protocols and supervise runs.
Open problems named in the post include verifying multi-step protocols and monitoring agent reasoning and execution traces. If you are building agents that act on real systems, the same layered pattern (simulate, validate, gate, log) applies, and AgentBeam, the agent security platform from the explainx.ai team, targets the "stop dangerous actions before they happen" step for software agents.
How it compares with other agentic-science work
Autonomous labs are now a crowded race. Anthropic described a Bay Area wet lab for biology and a life-sciences verification program. DeepMind's Antigravity work on an antimicrobial peptide used agentic science on a different biological target, and the institute's survey of bending the curve of discovery gives the wider frame. Sakana's AI Scientist v3 pursues the software-side version of the same loop, and Anthropic's Claude Science UV sky map shows the data-pipeline side.
What distinguishes Google's post is the focus on infrastructure: the digital twin and the validation step, rather than a single discovery. That is probably the more transferable contribution.
What this means if you build agents
- Put a simulator between the model and the actuator. VERID's dry run is the pattern to copy for any tool that has real-world side effects.
- Log everything, synchronized. Video plus action logs turned debugging from guesswork into diagnosis in the dye-stratification case.
- Watch for reward hacking. The myoglobin dip shows an agent will chase the metric you give it; correct with counter-screens like bead subtraction.
- Keep humans on the authorization step. Every vendor here keeps that in the loop, and it is the part regulators will ask about.
Related reading
- Anthropic's Bay Area wet-lab biology effort
- Anthropic life-sciences verification program
- DeepMind Antigravity and an antimicrobial peptide
- DeepMind Institute: bending the curve of discovery
- Sakana AI Scientist v3
- Claude Science and the UV sky map
- Official: Google Research post and PyLabRobot docs
Numbers reflect Google Research's October 16, 2026 post and its linked papers, and may change as the work is peer reviewed.