What happens to an agent’s reasoning when the environment stops answering instantly?
That question has been mostly theoretical for the past few years. Agents live in text, code, and browsers. Feedback arrives in milliseconds. A failed API call returns an error you can retry immediately. Now Anthropic has set up a lab in the San Francisco Bay Area for physical biology experiments, with a focus on rare diseases, and Claude can operate lab equipment. The feedback loop just got a lot slower and a lot more expensive.
I want to sit with that shift, because I think the interesting story here is architectural, not pharmaceutical.
Wet labs break most of what agents assume
Almost every agent design pattern in production today rests on assumptions that a physical lab quietly violates.
- Cheap retries. Agents are built to fail forward. Try, read the error, adjust, try again. A failed experiment consumes reagents, time on a machine, and possibly a sample that cannot be replaced.
- Fast observability. Text environments return state immediately. Biology returns state after incubation, after growth, after a run completes. The gap between action and observation stretches from milliseconds to hours or days.
- Clean state. Software agents can often reset. Contaminate a plate and you cannot undo it. The environment carries history that no one logged.
- Full observability of the action’s effect. A shell command either ran or it did not. A liquid handler can execute a command perfectly and still produce garbage because a tip was clogged. The action succeeded; the intent failed.
An agent operating lab equipment has to plan under all four constraints at once. That is a different design problem from anything a coding agent faces, and I do not think the solutions transfer neatly.
Long-horizon planning becomes mandatory, not optional
In software, you can get surprisingly far with a greedy agent that takes one step, looks at the result, and picks the next step. The loop is so cheap that shallow planning plus many iterations beats careful reasoning.
Physical experiments invert the economics. If every observation costs a day, the value of thinking harder before acting goes up sharply. The agent needs to reason about which experiment produces the most information, not just which experiment is next. That is closer to active learning and experimental design than to the tool-calling loops most agent frameworks implement.
It also forces something agents are famously bad at: holding a hypothesis across a long gap. Between issuing a command and reading a result, the agent has to preserve why it ran that experiment, what outcomes would confirm or disconfirm the hypothesis, and what it planned to do in each case. Write that state down badly and you get an agent that reads a valid result and has no idea what it means.
Why rare diseases is a revealing choice
Anthropic’s lab focuses on rare diseases, and the company has said it is not exclusively for drug discovery. I read that framing as a hint about what is actually being built.
Rare diseases are a data-poor setting. There is less prior literature, fewer prior experiments, thinner datasets. That is precisely where pattern-matching over a large training corpus stops carrying you and where the ability to run new experiments starts to matter. If you wanted to test whether a model can generate and test hypotheses rather than recall them, a data-poor domain is where the test is honest.
The “not exclusively drug discovery” part suggests the lab is also infrastructure for studying the agent itself. You cannot evaluate an agent’s scientific reasoning against text benchmarks alone, because the answers are in the training data. A physical lab produces ground truth that nobody has written down yet. That is an evaluation asset as much as a research one.
What I would want to know next
The public facts are thin, so I will be clear about what remains unanswered rather than guess at it. The details that would tell us the most:
- Where the human sits in the loop. Does the agent propose and a scientist approves, or does it execute plans directly?
- How experimental results are fed back. Structured instrument output, or images and free text the model has to interpret?
- Whether the agent maintains a persistent experimental record across runs, and what that memory looks like.
- How failures are attributed. Bad hypothesis, bad protocol, or bad hardware execution are three very different errors that look identical from a null result.
Those four answers would tell us more about the state of agent architecture than most benchmark releases. An agent that can run a coherent multi-week experimental program, and correctly diagnose its own failures along the way, has solved problems that the text-based agent world has been able to sidestep.
The pipette is not the interesting part. The clock is.
🕒 Published: