The Agent Had the Same Laboratory. What Changed Was the Evidence.

What they actually compared

Andromeda 2 is an agentic formulation designer built at Intrepid Labs in Toronto. It reasons over structured in-house experimental evidence, proposes and critiques candidate compositions, and hands executable batches to a miniaturized design–build–test platform. A deterministic feasibility engine checks compositional and robotic constraints before anything is prepared. A formulation scientist reviewed each proposed batch for safety and executability, but did not rewrite the compositions. ^[1,2]

The test case was paclitaxel in a self-emulsifying drug delivery system: a clinically important, poorly soluble compound whose oral formulations have to keep the drug apparently solubilized after gastrointestinal dilution. Each arm received a matched budget of 96 unique formulations, run as six batches of 16, on the same robots, operators, timing and dissolution protocol. The primary endpoint was the area under the apparent solubilized-concentration curve in fasted-state simulated intestinal fluid from 10 to 240 minutes. That is a physicochemical assay, not oral bioavailability. The authors say so. ^[1]

Against that endpoint, Andromeda 2 reached a 50% high-AUC hit rate. Andromeda 1, a sequential probabilistic optimizer already used across dozens of live development programs, reached 17%. The experimental DoE campaign reached 2%. Twelve Andromeda 2 formulations met all four target-product-profile objectives; Andromeda 1 found six; DoE found none. Median AUC was 70.1, 12.0 and 3.5 mg·min/mL. Maximum AUC was almost identical for the two Andromeda systems. ^[1]

That last number is worth sitting with. The agentic system did not find a higher isolated optimum. It produced more usable formulations inside a fixed wet-lab budget. The advantage also survived inside the productive region of composition space: among Type IIIB and Type IV lipid formulations, Andromeda 2’s hit rate was 49%, against 29% for Andromeda 1. Finding the right neighborhood was not the whole story. ^[1]

The historical evidence used by Andromeda 2 contained no paclitaxel formulations. The transfer being tested was from other APIs and other campaigns to a new one, then sequential adaptation to the measurements that campaign produced. ^[1]

The ablation is the result that matters

To isolate the contribution of that record, the authors ran Andromeda 2 again with historical in-house experimental evidence withheld. Everything else stayed the same: objective, design space, constraints, batch size, wet-lab workflow. After each batch, both configurations still received the paclitaxel measurements just generated. ^[1]

Campaign-wide mean AUC was 62.0 mg·min/mL with the record and 46.2 without it. High-AUC hit rate fell from 50% to 31%. The first full-TPP formulation appeared at formulation 12 with the record and at formulation 53 without it. ^[1]

The batch trajectories are more informative than the averages. The reduced-evidence system matched or outperformed the full-evidence system early on, then regressed. The full-evidence system improved across successive batches and held the gain. The authors’ reading is that historical evidence helped the system interpret new measurements, not merely seed a stronger first batch. ^[1]

That is a more specific claim than “data helps.” It says the laboratory’s earlier work is part of how the next result should be read.

We have already made a version of this argument for agents that reason over public biomedical databases and for computational campaigns that notice anomalies in primary sequence. Those papers showed that agents can work when the evidence is queryable. This one measures what happens, in a wet lab, when the organization’s own evidence is removed from an otherwise identical loop.

Most laboratories could not run this experiment

The ablation only works because Intrepid already had something to withhold.

The in-house record included prior compositions, measured dissolution, dispersion, droplet size, polydispersity, excipient behavior and platform-specific measurement history. New results were captured in the same structure. That is not a folder of PDFs. It is experimental memory that an agent can query while proposing the next batch. ^[1,2]

Most formulation and discovery groups do not have that object. The successful series sits in an ELN. The failed composition sits in a slide that never names the assay version. The reason an excipient was retired sits with the person who ran the program. You cannot ablate a record that was never assembled.

A closed loop that forgets its own history has to relearn the landscape on every new API. Andromeda 1 is built that way on purpose: each campaign starts largely de novo and learns from the measurements it generates. Andromeda 2’s advantage, on this evidence, is that it does not have to. ^[1]

The same requirement showed up in a different closed loop. In July, Chalmers, Gothenburg and Cambridge described an LLM-guided yeast platform that generates hypotheses, designs interventions, and runs automated cultivation and metabolomics. To make that loop executable they encoded hypotheses in description logics, extended a community phenotype ontology, and stored hypotheses, experiments, LLM traces and results in a graph database. One later hypothesis could be tested against data already acquired. Chalmers published a plain-language account of the work last week. The biology is modest: a handful of nutrient–stressor interactions in baker’s yeast. The infrastructure point is not. The loop depends on a machine-readable record of what was asked and what was already done. ^[3,4]

This is the same fragmentation problem we described for bioinformatics analyses. Results survive. The relationships that would let the next experiment use them often do not.

What has to exist before the next campaign can use the last one

If closed-loop formulation, or any other autonomous experimental campaign, is going to compound rather than restart, a few ordinary capabilities have to sit underneath the robot.

• A queryable archive of prior experiments: compositions or constructs, assay versions, measured outcomes, discarded candidates, and the reason a series was retired.
• Provenance that ties a new batch back to the claim it was meant to test, including the dataset or historical analogue the agent used to propose it.
• A trail complete enough that a scientist can inspect why one region of design space was favored and another was not, especially when the same campaign will not replay.
• Scoped permissions. An in-house formulation record is a different governance problem from a public metabolomics dataset.

The FAIR principles were written for reuse by humans and machines. Closed-loop laboratories make machine-actionability a campaign requirement rather than a repository afterthought. ^[5]

This is also one of the challenges we are exploring at OVAITY. We are building a Scientific Intelligence Platform that connects projects, experiments, datasets, analyses and decisions in one governed environment, so scientific context remains usable over time. Context-aware AI over that project record is being validated with researchers. Broader research agents and autonomous agent workflows remain prototypes. We are not building a formulation robot. We are trying to make the underlying scientific record coherent enough that a campaign, the evidence it reused, and the experiments that followed can be inspected as one piece of work.

Treat the laboratory record as part of the loop

The preprint has limits that should stay attached to it. The FaSSIF assay does not measure digestion, permeability or in vivo exposure. Each strategy was a single campaign, so the authors treat the statistical comparisons as descriptive. The DoE comparator explored the shared composition space directly, without the expert pre-screens that often accompany conventional formulation work. Transfer across chemically distinct APIs still has to be shown. ^[1]

None of that removes the operational finding. Give an agent the same laboratory and take away the structured experimental record, and the later batches get worse. Leave the record in place, and the system uses new measurements more productively.

A workbench can help a scientist design the next experiment. A robot can execute it. Neither replaces a durable account of what the organization has already tried.

Until that account exists, adding agents to a closed loop mainly accelerates the consumption of the current campaign.

If your team is exploring AI-native research workflows but struggling with context, traceability, or project memory, we’d love to learn from you.