November 21, 2026
Heard in AI

Leash Bio's premeasured lab trades agent freedom for replays

Ian Quigley of Leash Bio says its premeasured chemistry environments let AI agents replay a lab almost endlessly, but only with molecules already measured. He says Leash does not yet know whether the learning carries over to new biology.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on Latent Space, episode published October 26, 2026

Ian Quigley of Leash Bio has a short phrase for what his company builds for AI agents: "a lab in the loop with amortized latency." He used it near the end of a follow-up conversation on the Latent Space podcast, published October 26, 2026. In that conversation he described what the approach gives up, how fast Leash is producing these environments, and the question the company cannot answer yet.

The task the agents face

Leash builds test worlds for AI agents doing early drug discovery. An agent, such as a language model, gets a protein target and a catalog of molecules, plus "a budget of 100 molecules," and has to find the best binder it can, Quigley said. A binder is a molecule that sticks to the protein, a first step toward a drug. As it searches, the agent writes out hypotheses in plain language about why some molecules work and others do not, then spends its limited budget testing them.

Leash has already measured how every molecule in the catalog binds, so nothing is synthesized or tested while the agent works. "Because it's all pre-measured, you can roll out a lot of these," Quigley said. A rollout is one complete attempt by an agent. Leash was up to about 90,000 of them, he said. In its October 14, 2026 announcement of the environment, called Minos, Leash says agents test molecules in batches of at most five.

Recording once, replaying many times

Asked whether Leash had plans for other problems of the same kind, Quigley gave the organizing idea behind the project. In "lab in the loop" work, an AI picks experiments, a real laboratory runs them, and the results guide the next choice. Leash records the lab's answers once, then lets agents run "effectively infinite rollouts" through that recording. The company is thinking hard about other biological or scientific questions that could be turned into environments of this kind, he said.

He gave two reasons conventional lab-in-the-loop work is a poor fit for training agents. It "does take a long time," because each round waits on real experiments. Real experiments also have limited budgets, so the loop is "not counterfactually dense": it shows only what happened for the choices actually made, never what would have happened had the agent picked differently. To make agents learn, Quigley said, "you got to keep running it and do different choices with every rollout." Reinforcement learning, in which a system improves through repeated tries and feedback, depends on that kind of variation.

The conversation turned to how Leash's multiplexed screening, which measures many molecules against a protein at once, spreads the lab's cost over every later replay. Quigley accepted that and tightened his own phrase: "it's not quite accurate to say it's a lab in the loop with zero latency, but it's pretty close."

What a live loop can do that a replay cannot

The discussion raised Lila as a comparison, with the qualification that it is not the same thing: Lila also runs experiments and uses them in its rollouts, but not live. Lila's September 25, 2026 account describes a closed loop for green-hydrogen catalysts. Bayesian models and language models choose recipes, the lab synthesizes 96 catalysts in parallel, checks their composition, tests acidic oxygen-evolution activity and stability, and characterizes how they degrade, and those results shape the next batch. Lila says its scientists at first reviewed proposals extensively and later concentrated on safety checks, while staff still transferred samples between instruments.

The two setups differ in structure. Lila performs new physical experiments between rounds of choices. Leash replays an agent's choices against measurements it has already taken. Quigley stated the trade-off in both directions. Working "in the rollout one step at a time offers a lot more freedom because you can choose almost anything," he said. "And here you're restricted to the number of choices that we made when we pre-measured the environment." That is "definitely true," he said, before adding: "it's still a pretty big number of choices."

The video game problem

The same restriction had come up earlier, when the discussion turned to whether the environment could become a game people play, a kind of medicinal-chemistry video game. "I like this plan," Quigley said. Leash had made "a game a little bit like this" for Name That Chemist, a quiz introduced in a December 17, 2025 post that asks players to recognize which chemist a molecule came from.

Human players would meet the premeasured catalog immediately. A medicinal chemist, Quigley said, arrives with a personal suite of moves: "put a methyl here, like move this around." A premeasured environment can answer only for what Leash actually measured, so the game would have to tell the chemist, in effect, "you can't do whatever you want." The exchange landed on the design problem: players have to know what the available options even are. Quigley agreed that Leash would have to surface them "in a meaningful way," and said the game could be shrunk down to a smaller number of building blocks.

From two environments a day to twenty

Since starting Leash, the company has screened about a thousand proteins against a couple of chemical libraries, Quigley said. It now builds two new environments a day and is aiming for about 20, a target he placed at "the beginning of maybe 26 or 27." He hedged it: "I think we could probably get to 20 a day without too much lifting. So we'll, we're pushing. We'll see if we can do it." The Minos announcement gives the same two-a-day rate and 20-a-day plan. It also names possible next tasks: selectivity across targets, which asks whether a molecule binds one protein and not others, binding to mutated proteins, and other kinds of biological measurement.

Quigley presented his reason as an informal read of the field rather than a result. "Kind of informally, it seemed like where a lot of these agents are making the best and the fastest progress is where there's instant feedback that's very scalable," he said.

The question Leash cannot answer yet

That kind of feedback is "really hard to do in the physical world," Quigley said, but he thinks "a handful of biological questions" could be rebuilt this way. Then the caveat: "we don't yet know if we're able to take reinforcement learning progress from these kinds of environments to move on to new biology yet." He added: "we're going to find out. It's not that long."

One possibility raised in the discussion was that training on a covering set of targets might let agents generalize to cases in between. Quigley agreed and widened the question. He said he needs to talk to "more decision people," because much of the work is about how you spend a limited experimental budget and what you can do to get better at that, "beyond just medicinal chemistry." He put the skill as: "When do I stay? When do I go? How do I, how do I get better at that?"

Leash is also "incorporating more and more structural tool calls" into the environments, Quigley said. He called it "a pretty fun time right now."

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03

Connected ideas and articles

From the conversation

Podcast episodes

Latent Space

⚡ An Eval for AI Scientists That Knows the Answer Key - Ian Quigley, Leash Bio

Episode published This article draws on 2:11–3:05, 6:17–7:16, 12:31–16:12, 16:13–16:16 and 16:17–16:20 (approximate times)

Article history

Updates to this article

Tags