Here is the game Leash Bio sets for its AI agents. The agent sees 10 molecules that stick to a target protein and 10 close relatives that don't. It has to explain why the first group binds. Then it gets a catalog of millions of molecules and is told to test its idea by picking five or ten of them. The results come back immediately, "like it's a coding sandbox," Ian Quigley of Leash Bio said on the Latent Space podcast. The agent revises its explanation and picks again. Quigley said the goal is "to get the best binder in as few steps as possible very much like a medicinal chemist would do."
The answers arrive instantly because no new experiment is being run. Leash has already measured how those molecules behave. The agent is really querying a database of past experiments, but from the agent's side it looks like working in a lab.
The data behind the game
Leash works with DNA-encoded libraries, or DELs. These are huge pooled collections of small, drug-like molecules, and each molecule carries a DNA barcode. A protein is exposed to the whole pool at once. Sequencing the barcodes of the molecules that stuck shows which ones bound. Leash screens the same libraries against many proteins. It has also built Hermes, a model trained only on Leash's own data.
The episode was published October 14, 2026. According to the publisher's description, it was recorded about two months earlier, while the agent project was still something the team was "just playing with."
Quigley said the project began when Leash noticed it had a growing collection of two things. The first was "activity cliffs": pairs of nearly identical molecules whose binding behavior differs sharply. The second was profiles of the same chemicals measured across large numbers of proteins. The team wanted to know if reasoning agents could explore that collection in a "lab in a loop" setup. In that setup, a system chooses an experiment, sees the result and chooses the next one.
To get there, Leash took open-source language models, which Quigley said included Qwen and GLM, and taught them medicinal chemistry in plain English. The models learned to describe a structure in words and to give "the diff," or difference, between two molecules. Next they were shown binders and non-binders and asked why the binders bind. That step included "hard negatives": molecules that look like they should bind but don't. The models write their hypotheses in readable English, "which we think is great," Quigley said.
When Quigley described the lookup, the conversation framed it as creating a reinforcement rollout. Reinforcement learning trains a model through trial, error and reward, and a rollout is one attempt played out from start to finish. Quigley agreed, describing a source of rollouts that is "growing all the time," full of activity cliffs and protein variants. Because Leash knows that small changes along an activity cliff cause a reproducible, measurable change in behavior, he said, the agents' trajectories can also produce reasoning traces at a useful scale.
He compared this with a medicinal chemist's career. Chemists improve early drug candidates through structure-activity relationship (SAR) campaigns: they change a molecule, see how its behavior changes, and repeat. Quigley said a chemist might see "a couple dozen of these maybe 50" in a career. "Last weekend we pulled out over 10,000 of them and we're not even trying that hard," he said.
In the version he described, the agents got 10 binders and a budget of 100 molecules, "and they'll get the best binders out of six and a half million in that budget." His verdict was that "it seems like it's working." He noted one limit: the agents are confined to chemistry Leash has already screened. Later, he said, models could choose additional molecules from catalogs, which would then be ordered and retested, as in a traditional lab-in-the-loop approach.
Why not just scale Hermes?
The conversation turned to why this should beat a larger version of Hermes, an attention-based model. Quigley's first answer was memorization. Leash is happy with what Hermes attends to, he said. But as protein mutants are added to the training data, "as you add more parameters to a model like they're going to memorize the mutants too." Leash wants a system that understands the causal changes, and "using natural language it's like a pretty good regularizer for the problem." A regularizer is something that pushes a model to learn general patterns rather than fit the quirks of its training data.
His second answer was speed. Quigley said he knew of no other reinforcement-learning loop that is physically grounded and also this scalable and fast. Labs that use robots to run experiments in a loop still have limited resources, he said. They still have to decide what to make in the first place, and each loop takes longer. If you can measure a chemical landscape at scale, keep enlarging it and let agents play in it, "we think it's probably going to go a lot faster for this particular problem." The discussion also mentioned other groups pushing lab automation, with the point that they had not, as far as the questioner knew, turned pooled screens into a database lookup that supports huge numbers of rollouts.
The questioning pressed on why the approach should be effective rather than merely safe: it might avoid bad mistakes without getting things especially right. Quigley pointed to the rapid reward signal, like a coding sandbox. He said the key condition is that the data must include "the choices that a human being would not have made," which is possible because Leash measures all its molecules across all its proteins. The agent has to be able to make the counterintuitive pick and get feedback on it. "It's got to have the counterfactuals it's got to make mistakes," he said. Those mistakes, including possible future ones, must be present in the data without bias. Otherwise the agent learns "chemist sociology fingerprints." That phrase refers to another Leash finding discussed in the episode: models trained on public chemistry data can pick up who made a molecule instead of what makes it work.
Brandon Anderson, one of the hosts, summed the idea up this way. Through its language pretraining, the model already has an intuition for some basic chemistry, the kind most medicinal chemists might have. "And now you're giving it sort of like an infinitely fast lab in the loop." Quigley said, "that's correct." The discussion added that reasoning skills might matter even more: lessons from one campaign could help an agent reason about the next, harder question.
Using text as the input has a precedent that came up while the idea was being introduced. GenePT, from Stanford researchers, represents genes with text embeddings of their written descriptions. Embeddings are numerical summaries that a model can compute with. As the project's documentation explains, it builds these representations from NCBI gene summaries using a general-purpose text model, so accumulated human knowledge about each gene feeds into the representation.
Quigley stressed that Leash is small. "We're a tiny shop right there's only 10 of us," he said. He is looking forward to the experimental corpus growing at "2x 10x you know like 100x" its current rate, because the space for agents to explore would expand dramatically.
The Kosuri paradox
Anderson raised what he called his favorite question, the "Kosuri paradox" of machine learning in drug discovery. It is named after a tweet by Kosuri, whose own episode Anderson said he hoped to release in the coming weeks. The paradox runs like this: to build a good model you need lots of high-quality data, but if you can collect that data, why do you need the model? You could simply collect the answer.
Quigley partly agreed. If the only goal is good chemical starting material for one target, he said, the paradox holds: screen a lot and you already have the molecule. But he argued that chemical data spread broadly across many targets could produce a deeper machine understanding of chemical behavior. With that understanding, optimizing early hits could go much faster. He added that Leash has screened an enormous amount without raising much money.
The larger reason for the language-model work, he said, is a hope that strong chemical reasoning could reach well beyond finding new molecules. A system that is good enough might move straight into agriculture, veterinary science or materials science without screening all those proteins first. He compared this with protein-structure models, which now help teams find starting material for targets that are hard to screen. Quigley framed it as an aspiration: "we are hopeful" that enough engineered activity-cliff variation could get to a point where "you don't need to screen it all ever again."
What Leash released
On October 14, 2026, the day the episode was published, Quigley's team announced Minos. It describes Minos as replayable environments, each pairing a protein with a premeasured chemical library. The released setup differs from the one in the interview. The flagship task gives the agent no starting hit at all, rather than 10 binders and 10 non-binders. The agent tests 100 molecules in batches of up to five over 20 rounds, aiming to maximize binding enrichment, a measure of how strongly molecules are picked up in the screen.
The announcement's illustrated evaluation covers 33 targets and about six million molecules, with 10 trials per target. Leash reports that stronger models discovered modestly more and were more willing to leave unpromising regions of chemistry. It also explains the limits of replay. Measuring everything in advance supplies outcomes for the molecules an agent didn't choose, but only inside the screened library. Anything outside it still needs a physical experiment. The tasks measure the search for binders against a single target. They do not cover ADMET, the absorption, metabolism, toxicity and related properties a real drug must also get right.
The post reports roughly 100,000 independent search trajectories over the preceding six weeks. That is a different count from the 10,000-plus SAR campaigns Quigley described in the interview. Leash says it is releasing the environments and reasoning traces publicly. The publisher's notes add that a preprint and an open dataset accompany the announcement. The failures Leash reports are recognizable ones: agents left budget unused, tested the same molecule twice, and updated poorly after results contradicted their hypotheses.