The Autonomous Experiment Loop

Step one

Start with a question that can become a test

A useful autonomous run begins with a question narrow enough to become evidence. "Find a better catalyst" is a domain goal. "Compare these three catalyst families under this temperature window against this metric" is closer to an agent-ready task.

The model should restate the question, list assumptions, identify the evidence it needs, and ask for missing constraints before proposing experiments. Refusing to proceed can be a high-quality scientific act when the question is under-specified.

Step two

Generate hypotheses with explicit reasons

Hypothesis generation should not be a brainstorm dump. Each candidate needs a rationale, expected observation, contrary observation, and cost estimate. A multi-agent co-scientist pattern can help here by separating generation, ranking, debate, and critique.

The output should be easy to reject. A hypothesis that cannot be falsified by any available measurement is not ready for an autonomous run.

"structured scientific thinking engine"

Towards an AI co-scientist, Nature

Step three

Write the protocol before touching tools

The protocol is the contract between the agent and the reviewer. It should describe variables, controls, inputs, outputs, allowed tools, stop conditions, and review gates. For code experiments, pin the environment and include tests. For lab experiments, include safety constraints and physical limits.

An agent should not improvise a new protocol after seeing favorable intermediate results without marking that as a protocol change. Otherwise the system quietly turns exploration into confirmation.

Step four

Execute inside a bounded environment

Execution can mean running code, querying a database, launching a simulation, sending a robot instruction, or operating a browser. The environment should have budget limits, rate limits, data boundaries, and emergency stops. The agent should know exactly what it is allowed to do and what requires approval.

For physical systems, execution authority should be layered. A dry-run plan and simulator pass should come before instrument control. Human sign-off should be recorded as part of the run, not handled out-of-band.

"closed-loop experimentation"

Self-driving laboratories for chemistry and materials science, Nature Communications

Step five

Analyze results and preserve the uncertainty

The analysis step should separate raw observations from interpretation. It should report failed tool calls, rejected outliers, missing data, and alternative explanations. It should avoid rewriting the run as if the result was obvious from the beginning.

The best autonomous-science reports feel more like lab notebooks than press releases. They make it easy to see where the agent was right, where it guessed, and where the evidence remains weak.

Step six

Critique before the next run

The critique step should ask whether the result is novel, whether controls were sufficient, whether the method was reproducible, and whether a cheaper disconfirming test exists. It should include both self-critique and an independent review path where possible.

Only then should the agent propose the next experiment. Closed-loop autonomy is valuable because it can learn from measurements, but it is dangerous when critique is shallow.