Google Research is connecting AI agents to laboratory robots, experimental readouts and molecular design loops. The most important part is not that the machines can follow instructions. It is that they can be corrected by physical reality.
October 2026 | AI Agents | Biology | Autonomous Laboratories | Google Research
For years, the phrase AI for science has usually described something that happens on a screen. A model predicts a protein structure, searches an enormous dataset, proposes a molecule or helps a scientist make sense of results that already exist.
Google Research is now pushing that boundary into the laboratory itself.
On October 16, 2026, researchers at Google described three connected projects intended to move AI agents closer to a closed scientific loop: propose an experiment, test it in the physical world, observe what actually happened, revise the hypothesis and try again.
The long term idea sounds almost like a science fiction laboratory. An AI system could explore an enormous biological search space while robotic equipment performs the experiments needed to tell it when it is wrong. But the work Google has published is more careful, and more interesting, than the phrase autonomous laboratory might suggest.
These are not unsupervised machines discovering medicines on their own. Humans still set scientific goals, authorize designed sequences, verify protocols and supervise robot runs. The three projects are also not yet one seamless autonomous system. Google describes them as building blocks that it now wants to weave together.
That distinction matters, because the real story is not that an AI has suddenly become a laboratory scientist. It is that researchers are beginning to build the infrastructure that could let an AI participate in the entire experimental cycle while remaining constrained by measurement, simulation and human oversight.
The difficult part is not having an idea. It is touching reality
Biology is a difficult place for purely computational intelligence. The space of possible molecules is vast. Google gives one simple example: a protein only 100 amino acids long can, in principle, occupy at least 20 to the power of 100 possible sequences. No scientist, computer or AI system can simply inspect every possibility.
That makes search strategy extremely important. Models can generate candidates quickly, but biology refuses to stay inside the model. Molecules fold in unexpected ways. Reagents behave differently depending on temperature, timing and concentration. Pipette tips carry tiny traces of liquid. Instruments fail. A design that looks elegant in simulation can behave badly when it reaches a real assay.
The laboratory is therefore not merely where an AI confirms its answer. It is where the AI discovers that its answer may have been wrong.
This creates a second problem. Frontier models are powerful at language and reasoning, but they are still unreliable at the kind of three dimensional physical state tracking that laboratory robotics requires. Code can be syntactically perfect and physically impossible. A model might ask an eight channel pipette to enter a container that is too narrow, try to aspirate from a well that is already empty, or execute a protocol that is mechanically valid but scientifically distorted by a fluid effect it did not model.
Google’s first project is designed around that gap between a convincing plan and the stubborn physical world.
VERID gives the agent a live model of the laboratory
Google calls the framework VERID, short for VERified Infrastructure for Discovery Laboratories. It is essentially a semantic digital twin of the laboratory workcell: a live three dimensional representation that keeps track of the robot, the equipment, the geometry of the deck, the volume remaining in individual wells, the identities of reagents and even the available pipette tips.
The system is built on the open source PyLabRobot framework. Before an AI generated protocol reaches the physical liquid handling robot, VERID can simulate the proposed procedure in seconds. The agent can inspect the current state of the laboratory, perform a virtual dry run, receive warnings about impossible actions and replan before any reagent is touched.
Google also records synchronized video and action logs. That turns the experiment into something closer to an auditable computation. The researchers can see not only what result appeared at the end, but what the robot did, what the agent believed the state of the lab to be, which variables changed and where an error entered the process.
Several of Google’s case studies show why this matters. Video and execution traces revealed uneven dye stratification caused by dispensing too close to the bottom of a well, which allowed the agent to alter its mixing strategy. Persistent state tracking allowed the system to recover from a plate reader failure without repeating earlier pipetting steps. In another test, the simulator rejected an eight channel aspiration that physically would not fit into a narrow trough. The agent then replanned and designed a targeted factorial experiment that isolated a subtle bias caused by residual liquid on the pipette tip walls.
In a separate self play campaign, an agent repeatedly designed and ran experiments to reconstruct blinded dye mixtures from overhead camera images. Across repeated cycles it identified optical and fluidic artifacts in its own setup and improved accuracy by more than fivefold, according to Google.
There is an important boundary here. These VERID demonstrations used controlled test materials such as dyes and water. They should not be mistaken for an AI freely performing complex biological experiments. Their purpose was to test something more fundamental: whether an agent can remain grounded in the physical state of a laboratory instead of merely generating plausible robot code.
The aptamer experiment is where the story becomes biological
The second project used DNA aptamers as a faster testing ground for laboratory learning. Aptamers are short strands of DNA that can fold into shapes capable of binding to a particular molecular target. They are useful here because large pools of DNA sequences can be synthesized, tested and then read by sequencing, making the design, test and analysis cycle much faster than many other forms of molecular discovery.
Google reports that one complete cycle took about seven to ten days. The agents designed candidate sequences. Those sequences were synthesized and screened against proteins. Sequencing data from the surviving candidates then became evidence for the next round.
This sounds straightforward until the laboratory pushes back.
For alpha thrombin, the first round produced no confirmed binders. Google says fixed primer overhangs interfered with folding in 48 percent of sequences. The agents changed their design strategy by introducing clamped flanking stems, and the confirmed hit rate rose to 26.0 percent in the second round. By the third round it reached 42.1 percent, with 98.5 percent specificity.
The myoglobin campaign is even more revealing because the system initially became worse.
In the second round, the agents heavily favored G quadruplex motifs. Raw binding measurements appeared to improve, but the wet laboratory showed that many of those sequences were simply sticky. They bound not only to the intended target but also to off target proteins and even to the magnetic beads used in the assay. Target specificity fell from 75.0 percent to 19.1 percent, while the bead corrected hit rate dropped from 10.2 percent to 3.7 percent.
The experiment had caught the agent optimizing the wrong signal.
Using that physical evidence, the agents revised their design rules for the third round. They limited the promiscuous motifs, explicitly accounted for bead binding and shifted toward different structural scaffolds. Specificity recovered to 72.4 percent and the confirmed on target hit rate rose to 33.5 percent.
That failure is arguably more important than the final number. It demonstrates why an autonomous science system cannot be built from reasoning alone. If the agent had remained inside its computational world, the sticky sequences could have looked like progress. The laboratory revealed that the apparent success was an artifact, and the next design cycle changed because of it.
From DNA strands to protein binders
Google’s third project moved from aptamers to de novo protein binder design, a more computationally demanding problem.
The target was fibroblast activation protein alpha, or FAP, a cancer associated protein used in the research as a molecular target. Designing a new protein binder generally involves proposing a candidate structure, designing a compatible amino acid sequence and then predicting how the candidate and the target will fold together.
That final refolding step can consume most of the computational budget, especially when the target protein is large. Google’s team used a method it calls epitope excision: instead of repeatedly folding the entire target, the system focuses on the local region where the binder is intended to attach.
Google reports that this reduced the structure prediction cost by a factor of 8.6 without sacrificing structural accuracy in the reported tests. It also created a useful role for the agent. A human scientist could set the high level strategy while the agent adjusted tactical choices such as where to bind and how much surrounding structure to include.
At equal compute, Google says this human and agent collaboration nearly doubled the in silico hit rate compared with a scientist working alone and produced promising candidates across a broader range of target sites.
The candidates were then screened experimentally using yeast surface display and fluorescence activated cell sorting, methods that allow large libraries of engineered proteins to be tested in parallel. The best designs included binders with single digit nanomolar affinity for FAP.
That is strong molecular binding, but it is not a cancer treatment. A binder can attach tightly to a target and still be far from becoming a useful therapy. It may fail on selectivity, stability, delivery, toxicity, behavior in living tissue or many other requirements that exist between an interesting laboratory molecule and a medicine.
The result is therefore best understood as evidence that an agent assisted design workflow can produce experimentally validated molecular candidates, not evidence that AI has autonomously discovered a cancer drug.
So how autonomous is the autonomous wet lab?
The word autonomous deserves careful handling here.
Google’s current system still places humans at several critical points. Scientists authorize and review designed sequences. They verify wet lab protocols. They supervise robot runs. Every designed sequence is screened for biosecurity concerns. The biological testbeds were deliberately chosen to be controlled, noninfectious systems without gain of function risk.
VERID also acts as a gate between model reasoning and physical machinery. An agent does not simply write code and send it directly to a robot. The proposed action is simulated, checked against the current digital twin and blocked if it violates physical constraints.
The architecture therefore looks less like an unsupervised artificial scientist and more like a layered scientific control system. A human supplies goals and authority. The agent explores possibilities and proposes actions. A digital twin checks whether those actions make physical sense. Robots execute approved procedures. Instruments produce measurements. Those measurements return to the agent and to the human researchers.
That feedback loop is the important part.
A language model can always produce another explanation. A laboratory can tell it no.
Safety is not separate from the science
Connecting powerful AI systems to biological laboratories creates obvious safety questions. Google’s work is notable because the safety mechanisms are part of the experimental architecture rather than an afterthought added around it.
The researchers describe four layers that matter. They chose comparatively safe test systems. Designed sequences go through biosecurity screening. Robot protocols are simulated before execution. Humans retain authorization and supervision.
VERID adds another layer that may become increasingly important as scientific agents grow more capable: traceability. If an agent changes a protocol, chooses an unusual design or reacts to an unexpected measurement, the system can preserve the state, action logs and synchronized video needed to reconstruct what happened.
This matters because a scientific agent does not need to behave maliciously to create risk. It can simply optimize the wrong thing. The myoglobin experiment is a small and harmless example. The agent found sequences that appeared to bind well, but many were actually nonspecific sticky binders. The wet lab exposed the shortcut.
In a more complex biological system, the equivalent shortcut could be harder to notice. Future autonomous laboratories will therefore need ways to distinguish true scientific progress from experimental artifacts, proxy optimization and unexpected interactions with the physical setup.
Google itself describes these as open problems. Its October report explicitly frames fully autonomous discovery as a future direction, not something already solved.
The real change may be the shape of scientific work
The most interesting implication is not that robots will replace scientists at laboratory benches. It is that the division of scientific labor may change.
Human researchers are good at choosing meaningful questions, understanding biological context, judging whether an objective is worth pursuing and recognizing when a result matters beyond a benchmark. AI agents are increasingly good at exploring large spaces of possibilities, maintaining many candidate hypotheses and carrying out repetitive tactical work. Robots are good at executing physical procedures consistently. Instruments are good at measuring what actually happened.
Put those pieces together and the laboratory begins to look like a closed information system. Ideas can move from language to code, from code to a physical experiment, from an experiment to measurements, and from measurements back into the next generation of ideas.
That could change the speed of some kinds of discovery, but not because biology suddenly becomes instant. Google’s aptamer loops still took days. Physical experiments still require materials, preparation, measurement and time. What changes is the amount of reasoning and experimental planning that can happen continuously around those slow physical steps.
In other words, AI may not remove the laboratory bottleneck. It may make the laboratory the most valuable source of truth in the entire system.
What remains unproven
There are several reasons not to race ahead of the evidence.
The three technical reports were released by Google alongside its own research announcement and, at the time of publication, had not been independently peer reviewed. The experiments were controlled research prototypes. The VERID robotics demonstrations, the aptamer loop and the protein binder workflow are related pieces, but Google has not yet shown them operating as one general purpose autonomous biology laboratory.
The biological campaigns also cover a small number of targets. Success on myoglobin, alpha thrombin and FAP does not establish that the same approach will generalize across the much messier landscape of drug discovery, cellular biology or disease research.
The system still depends on human judgment at important points, which is a strength for safety but also means claims of full autonomy would be premature. Independent replication will matter. So will comparisons with strong human teams, conventional automation and other AI assisted discovery systems.
And there is a deeper question that may become more important than raw capability: how do we know when an agent has learned the biology rather than learned how to exploit the experiment? The sticky aptamer episode is a reminder that nature does not reward elegant reasoning. It rewards whatever actually happens in the assay.
When the experiment can answer back
There is something fundamentally different about an AI system whose ideas can collide with the physical world.
In ordinary conversation, a model can sound convincing while being wrong. In a laboratory, a pipette can fail to fit. A well can be empty. A molecule can refuse to fold. A supposedly excellent binder can stick to everything. Reality becomes part of the training signal.
Google’s work is still early, carefully constrained and far from the image of an autonomous scientist discovering medicines alone overnight. But it shows the outline of something important: AI agents that do not merely generate hypotheses, but submit those hypotheses to experiments that can contradict them.
The next step, Google says, is to combine VERID’s traceable laboratory automation with the closed loop learning demonstrated in molecular discovery. If that works reliably, the most significant development may not be the robot arm or the language model by itself. It will be the loop connecting them.
An AI proposes. The laboratory answers. The AI changes its mind.
That is much closer to science.
Sources
Google DeepMind and Isomorphic Labs, “Our approach to bioresilience,” July 16, 2026.
Google DeepMind, Frontier Safety Framework, Version 3.1.
Prepared for BunCosmos.ai | October 2026