What changes when an AI agent gets laboratory feedback

A laboratory result can tell an AI agent that its hypothesis was wrong. It can also tell the agent that its experiment was wrong. Distinguishing those possibilities is one of the most important tasks in autonomous science. A disappointing measurement might reflect a weak molecular candidate, a misleading assay, an incomplete mixing step or an instrument failure. More generated hypotheses do not solve that ambiguity.

Google Research’s October 16 overview connects three projects: laboratory infrastructure, aptamer discovery and protein binder design. They address different parts of a design–test–learn loop. The ambitious direction is understandable, but the evidence becomes clearer when each experiment’s physical system, human contribution and measured endpoint stays visible.

The most revealing infrastructure work is VERID, whose currently linked manuscript is dated October 22. Its case studies use dye and deionized water. This article explains and adapts that paper under its stated Creative Commons Attribution 4.0 licence, crediting Yves Delley and colleagues. Discussion of the other manuscripts is deliberately narrower, and the examples and implications identified as our analysis are not additional experiments.

The laboratory needs a memory of what physically happened

VERID extends the laboratory’s software representation beyond a script that starts, issues commands and exits. It maintains a persistent semantic digital twin: a structured representation of the workcell, its equipment, containers, locations, contents and operating state. A well belongs to a plate. A lid can obstruct access to that plate. A tip has a state, and a reagent has an identity and an estimated remaining volume.

This hierarchy gives an agent something more precise to query than a prose description of the bench. It can ask about a well’s volume or whether an item is accessible before proposing an operation. The distinction matters because the validity of a command depends on current conditions. A transfer that was valid earlier can become impossible after the source empties or the available tips are consumed.

The implementation builds on PyLabRobot, a Python interface for supported liquid handlers and other laboratory equipment. That existing layer provides hardware abstractions and atomic operations. VERID’s contribution is the longer-lived state, simulation gate and synchronized records surrounding modular command blocks. This is an infrastructure composition, rather than an LLM independently acquiring a robot’s full physical understanding.

A useful comparison comes from the Opentrons protocol reference. Its documentation distinguishes instruments loaded into a protocol context from everything physically installed on a robot. Similarly, a list of loaded labware is not an inventory of every object occupying the deck. These documented distinctions show why a program’s internal objects and the actual workcell should not casually be equated. They are not evidence of a flaw in either system.

Persistent state addresses continuity, but it also creates an obligation: the representation must remain synchronized with reality. If a person moves a plate without updating the system, a perfectly consistent software model can become physically wrong. The conceptual value of the twin lies in making state explicit and queryable; it does not remove the need for sensing, reconciliation and controlled human interaction.

A simulation gate catches the mistakes it models

Before dispatching a command block, VERID clones the live state and runs the proposed operations against that virtual copy. The simulator checks modeled geometry, volumes, resource inventories and other constraints. It can fast-forward through waiting periods rather than consume laboratory time. A rejected plan can therefore be revised before using a tip, moving hardware or spending a reagent.

One reported example is especially concrete. An eight-channel operation required more horizontal span than a narrow trough allowed. The simulator rejected the move, and the agent revised its use of channels. This was not a language formatting problem: a command could be perfectly valid Python while describing an impossible physical placement. The relevant evidence was the relationship between tool geometry and container geometry.

Other checks concern conservation and inventory. An empty source cannot supply a requested volume. An occupied or covered destination can be inaccessible. A protocol that consumes more tips than the deck contains cannot finish as planned. Modeling these constraints makes the boundary between a proposed experiment and an executable experiment explicit. It also avoids relying on the language model to remember every quantity consistently across a long conversation.

However, VERID’s simulator is described as a discrete accounting system, not an exhaustive simulation of fluid dynamics and optics. That limit is essential to interpreting its demonstrations. A plan can pass volume and geometry checks yet mix a solution poorly or produce a misleading camera image. Real measurements and traces remain necessary precisely because some important effects fall outside the model.

Our interpretation is that this division of labor is more credible than expecting one general model to predict everything. Deterministic checks handle constraints that can be represented explicitly. Measurements expose what those checks omit. The agent can then use the discrepancy to design a diagnostic experiment. The resulting system gains reliability by making its model incomplete in a visible way, rather than silently treating modeled feasibility as scientific correctness.

A dilution problem becomes a measurement problem

In one VERID case study, an agent attempted serial dilution of colored solutions. The expected pattern was a regular concentration reduction from one column to the next. Observations departed from that expectation and varied across replicates. The agent suggested insufficient mixing, while a human operator examined synchronized video and execution traces to identify the physical causes more precisely.

Dispensing and mixing near the well floor left dye stratified. Small channel-to-channel height differences also caused some tips to press against the bottom, partially obstructing their openings. Both effects could produce an apparent concentration problem. Neither was adequately described by saying the commanded liquid volumes were correct. Nominal volume and actual homogeneous transfer were different properties of the operation.

The researchers revised the protocol’s vertical positioning and mixing behavior. The paper reports that this brought the dilution pattern closer to the intended twofold progression and reduced discrepancies between replicates. These are results in the stated dye system. They show that linked measurements and visual records can support protocol improvement; they do not establish reliable execution of arbitrary biological assays.

The human role should remain part of the finding. The agent did not independently identify every root cause in this first demonstration. The operator used visual evidence to refine the explanation, after which the agent adjusted the protocol. Calling the entire episode fully autonomous would erase a contribution that was central to diagnosis and to understanding what the system demonstrated.

The broader analytical point is that a result needs provenance at the level of operations. A final table of absorbance values cannot reveal whether a tip contacted a well floor. Video without synchronized command timing may leave the relevant moment ambiguous. Putting those streams on a common timeline changes what an investigator can infer, because an unexpected measurement can be traced to the action and state that preceded it.

Recovery is not the same as replay

The revised dilution run encountered an instrument-state failure after liquid handling had already finished. VERID’s persistent state supported continuation without repeating the transfers. The prepared plate could be moved, the reader recalibrated and the measurement completed. This matters because repeating an entire script can be actively wrong when some physical actions have already occurred.

In ordinary software, retrying a failed request can be harmless if the operation is idempotent: repeating it has the same effect as doing it once. Many laboratory actions are not idempotent. Dispensing another volume into a well changes its contents. Repeating a dilution changes concentration again. Consuming another tip changes inventory. A recovery policy must know which actions completed and which remain uncertain.

Our original interpretation is that the useful abstraction is a partially completed physical transaction. A system needs evidence about its last known state, not merely an exception message. If a communication failure happens between movement and acknowledgement, the software may not know whether the movement occurred. That uncertainty should be represented and resolved rather than converted automatically into an instruction to retry.

The VERID demonstration supports a specific instance of stateful recovery. It should not be generalized into a guarantee that every interrupted action can be safely resumed. Different failures can invalidate different parts of the twin. Loss of an instrument reading, uncertain liquid transfer and an unexpected manual intervention require different reconciliation procedures. A recovery success is most informative when the interrupted step and retained evidence are both stated.

This also changes how an autonomous laboratory’s productivity should be measured. Counting completed planned runs misses time and materials spent recovering from partial failures. A system that resumes correctly may be more useful than one that generates a protocol faster but repeatedly discards usable work. The relevant economic unit is a trustworthy completed measurement, including recovery and review, rather than a generated script.

The most informative test changes the suspected cause

A second VERID case study encountered a channel-associated discrepancy. An obvious explanation would be that one channel delivered too little liquid. Simply repeating the same arrangement might have strengthened that impression. Instead, the diagnostic changed which channels participated in an earlier mixing operation. The discrepancy followed that assignment rather than remaining attached to the original channel.

The paper attributes the effect to dye carried on the inner walls of mixing tips and subsequently washed into the assay. In that setting, the apparently low channel reflected a cleaner baseline; other channels had received extra material. This reverses the naive interpretation of the measurement. A relative deficit can arise because the comparator is inflated, rather than because the supposedly deficient item is faulty.

That example makes a substantial claim about experimental design, not only robot control. An informative intervention changes the factor that distinguishes competing explanations. If the suspected cause is channel hardware, move the treatment while retaining channel identity. If the suspected cause is an earlier liquid-handling history, change that history. A useful agent must decide which comparison can separate those accounts within available resources.

The study also makes clear why more replication is not always the answer. Repeating a biased arrangement can estimate its biased mean very precisely. It does not necessarily identify the bias. A factorial diagnostic can be more valuable than additional identical runs when the task is to distinguish mechanisms. This is our methodological interpretation of the reported case, not evidence that agents generally outperform scientists at causal reasoning.

For a prospective laboratory deployment, the implication is to preserve interventions as first-class data. Which channel was used, whether its tip was previously wetted and which plate position received the material can become explanatory variables. A dataset containing only candidate identity and final score may discard exactly the information needed to recognize an assay artifact.

A blinded challenge keeps the score outside the solver

VERID’s self-play demonstration used blinded dye mixtures and limited physical resources. A solver could design measurements using camera observations, while a separate plate-reader measurement supplied an independent reference after the run. The paper compared single-cycle and multi-cycle approaches and examined the use of general procedural learnings across challenges. This arrangement separates information available to the solver from information used to evaluate it.

That separation is particularly valuable in agent research. If an agent can access the answer, an apparent improvement may reflect leakage rather than a better experimental strategy. Here, the reference readout is used to score the solver’s predictions rather than guide its within-run decisions. It provides a concrete boundary that readers can inspect when assessing claims about learning from physical feedback.

The camera task is not trivial merely because it uses dyes. Broad color channels, saturation, reflections, automatic white balance and fluid handling can complicate inference. The scientific interest lies in how a system responds to those measurement limitations while deciding what additional observation would be useful. The low-cost materials make that interaction easier to study without claiming that a dye challenge reproduces the full complexity of biology.

The resource cap also prevents a misleading definition of success. An unconstrained agent could keep measuring until the answer becomes obvious. A constrained challenge instead asks whether it allocates experiments intelligently. Those are different capabilities. Reporting the available tips, wells, observations and time is therefore part of defining the task, rather than incidental implementation detail.

A transferable learning summary is another limited but useful object. It can communicate general lessons about a measurement process without revealing the hidden composition of a later sample. That is different from training a new model’s weights, and different again from letting a solver inspect previous answer labels. The word learning should identify which of these mechanisms actually occurred.

Additional cycles can initially make a result worse

The detailed self-play trajectory is more informative than a single improvement multiplier. In one challenge, the multi-cycle solver’s additional measurements worsened its aggregate error. Concentrated additions saturated a camera channel, and small transfers introduced variability. The agent acquired more observations, but those observations did not automatically carry more usable information. The paper describes a later challenge that benefited from general operating bounds learned from that earlier failure.

The later strategy used a sparse survey and comparisons across dilution trajectories rather than relying on individual spiked wells. This is a concrete change in experimental strategy: measure how a signal behaves over conditions, instead of assuming one measurement provides an unambiguous signature. It also illustrates why a claim that feedback always improves results would be incorrect. Feedback can expose a bad strategy before it supplies a better one.

The camera itself was part of the inference problem. Oblique views could mix illuminated well walls into the region being measured. The paper describes using calibrated geometry to focus on well-floor regions and account for illumination. This is a useful example of combining symbolic physical information with visual analysis. The agent did not need to infer the entire workcell from pixels when spatial relationships were already available through the twin.

Different error sources also demanded different responses. Saturated image channels called for a change in measurement range. Unreliable small transfers called for operational bounds. Concentration gradients called for attention to preparation and mixing. Treating all three as generic noise would encourage more repetitions of a flawed measurement. The study’s strongest examples instead connect an observed residual to a testable physical explanation.

The interface separates proposing, verifying and dispatching

The VERID methods describe distinct tools for inspecting state, verifying a command and dispatching it. Verification runs against a forked replica; dispatch re-verifies against the live state before physical execution. That second check matters because a plan approved against one snapshot can become stale if the workcell changes before it runs. A prior valid simulation should not be treated as permission to act against any later state.

Execution returns a run identifier, and the agent can poll for completion before examining the resulting dossier. This creates an explicit boundary between issuing work and receiving its consequences. The dossier then places state snapshots, commands, video and assay outputs into the agent’s workspace. Those are much richer observations than a single success message, and they make it possible to connect a proposed change to an actual run.

The reported implementation stores discoveries as structured natural-language operating bounds and analytical models. It does not require interpreting every improvement as a change to neural-network weights. That distinction is important for reproducibility: a later run can benefit from an updated knowledge base even when its underlying model checkpoint is unchanged. A comparison that records only the model name would miss that source of improvement.

The paper identifies automated indexing and knowledge extraction across archived dossiers, and systematic testing of their downstream benefit, as future work. That leaves a clear next research question: can the detailed records that help diagnose these individual demonstrations reliably improve new campaigns at scale? The infrastructure supplies a plausible route to answering it; the present case studies do not complete that broader validation.

The molecule projects answer narrower, different questions

The aptamer manuscript studies three feedback rounds on two protein targets. It reports that designs could exploit misleading binding signals and then recover selectivity after counter-screening. Its single-clone measurements also distinguish literature-derived motifs from strictly new scaffolds. This is a bounded discovery study, not evidence of general autonomous drug development.

The epitope-excision manuscript, currently dated October 23, separates computational hit efficiency from yeast-display recovery. Human–agent collaboration leads the former comparison; other arms recover more exact designs in selection. The autonomy experiment concerns one FAP target, one backbone class and one LLM controller. A modified variant also changes the apparatus and cannot isolate autonomy alone.

Those distinctions explain why a single headline such as “better binders” is inadequate. Computational screening asks whether a design passes a prediction-based rule. A display-selection experiment asks whether constructs enrich under a particular assay. Subsequent characterization asks about measured binding under stated conditions. Each stage eliminates some failure modes while leaving others unresolved. They should be connected as a chain of evidence, rather than treated as interchangeable success counts.

The underlying tools also perform different jobs. RFdiffusion generates protein structures under specified conditions. ProteinMPNN designs or scores sequences in relation to a backbone and exposes controls for fixed residues, sampling and model choice. Calling both simply “AI” obscures the pipeline: a controller coordinates specialized components whose outputs remain proposals until tested.

Our analytical contribution is to view these stages as a sequence of changing denominators. A generation count is not the number synthesized. The number synthesized is not necessarily the number expressed successfully. The number tested is not automatically the number confirmed by a different assay. Comparing percentages without identifying the population behind each can create an apparently impressive gain that comes from selection rather than improved underlying performance.

An evidence ladder prevents accidental overstatement

This ladder changes how experimental results should be communicated. A system can advance one stage without advancing all of them. Better simulation can reduce invalid operations without changing molecule quality. Better candidate ranking can reduce screening cost without improving the best final lead. Better measurements can reveal that an earlier apparent success was an artifact. Each outcome may be valuable, but it is valuable for a different reason.

Consider a hypothetical campaign with a fixed budget for one hundred candidates. One approach generates many proposals and selects the highest computational scores. Another reserves part of the budget for controls and alternative measurement conditions. The second approach might report fewer initial hits while producing greater confidence that those hits reflect the intended interaction. A raw hit count cannot decide which campaign produced more useful knowledge.

The accounting should also distinguish human effort. If researchers choose targets, approve sequences, prepare instruments and interpret difficult failures, those hours belong in the workflow. Automating tactical decisions may still be a significant advance. It becomes easier to judge that advance when the publication identifies which decisions moved to the agent and which remained with experienced scientists.

No clinical conclusion follows from these infrastructure and screening demonstrations. A measured binder is a research result with a particular assay and target context. The additional work between a promising research molecule and a useful intervention is outside the claims established here. Keeping the article focused on the actual endpoints makes the technical progress clearer rather than diminishing it.

What a convincing next comparison would measure

A strong follow-up would separate improvements in the controller from improvements in its tools and experimental apparatus. If a newer workflow changes the scoring metric, library construction and agent policy simultaneously, a better endpoint cannot be assigned to autonomy alone. Matched comparisons need to hold relevant conditions fixed or clearly state which causal interpretation is unavailable.

Resource matching deserves the same care. Equal candidate counts can conceal different compute costs per candidate. Equal compute can produce different numbers of candidates. Equal numbers entering a screen can conceal different diversity or selection procedures. A defensible comparison should choose the resource relevant to its claim, report the others and explain the stage at which each population was selected.

The ideal evidence would include failures that do not resolve neatly. An agent that stops because the measurement is ambiguous may behave more responsibly than one that confidently supplies a wrong explanation. An experiment aborted by a simulation gate is different from a robot crash, and both differ from a scientifically uninformative completed run. Collapsing them into one failure rate prevents readers from understanding where the system needs work.

Our proposed reporting unit is therefore a traced decision: the observation available, the action selected, the constraint applied and the resulting measurement. It can show whether a controller learned a transferable rule, merely repaired one local problem or benefited from a human diagnosis. This is a reporting proposal, not a capability we have independently measured in the Google systems.

The operational foundation is already visible in the Opentrons API introduction: a protocol declares equipment and then commands physical actions. An agentic system adds decisions around that substrate, but it cannot escape the substrate’s geometry, resource limits and state. More fluent reasoning does not make those requirements less concrete.

The feedback loop is an information system

A laboratory feedback loop is often drawn as a simple circle: design, build, test, learn. The circle hides the work of deciding what a test actually means. VERID’s value is that it makes several links inspectable: proposed actions are checked against live state, operations are recorded with observations, and unexpected outcomes can trigger a diagnostic rather than another blind iteration.

The new work is most persuasive where it is most specific. A geometry check prevented an impossible movement. Persistent state enabled a particular recovery. A changed channel assignment helped identify a liquid-handling artifact. Those are meaningful technical findings without turning three projects into a single end-to-end demonstration of autonomous biological discovery.

For readers following the wider agent field, our explanations of agent harnesses, reasoning models and LLM observability offer complementary context. The laboratory adds a demanding constraint to each topic: actions alter a physical world that cannot always be reset, and a useful answer must remain connected to the evidence that produced it.