Key points
- The agent operated a liquid chromatography-mass spectrometry instrument during an oxidant screen for synthesizing an organic dye.
- Its oxidant ranking matched one from scripted peak integration, and organic chemists confirmed its spectral interpretations.
- In simulation, it measured the highest-yielding condition at about its seventh injection, versus 14.5 on average for random order.
- Every action that changes the hardware's state must be authorized, validated and auditable.
Researchers have built an AI agent that can run a laboratory instrument and decide what to measure next, while every action it takes on the hardware stays within limits that humans set. It is designed for self-driving laboratories, which are automated labs that plan, run and analyze experiments. The work appears in a preprint posted on arXiv and has not yet been peer-reviewed.
The team argues that sharp data analysis alone does not make a lab autonomous. An agent also has to be trusted to operate equipment without supervision, and it has to know when to hand a decision to a person. Their design places the agent at the handoff between analyzing results and designing the next experiment. That handoff links two stages of the Design-Make-Test-Analyze (DMTA) cycle. Any operation that changes the state of the hardware must be authorized, validated and auditable.
The researchers tested the system on a liquid chromatography-mass spectrometry (LC-MS) instrument during a screen of oxidants for making an organic dye. The agent submitted measurement jobs and read mass spectrometry and diode-array detection data. It also flagged failed experiments and ranked the oxidants. Its ranking agreed with one produced by scripted peak integration, and organic chemists confirmed its spectral readings and failure calls.
In a simulation that replayed the archived data, the agent chose which wells to inject. It reached the highest-yielding condition at about its seventh injection, compared with 14.5 on average when wells were picked in random order. It found that condition in every campaign that allowed eight or more injections. The authors describe this as a step toward greater autonomy for self-driving labs.
Why it matters
The authors say the work shows that an agent can link analysis to experiment design while its actions stay bounded and verifiable. They see that as a step toward more autonomous self-driving labs.
What we don't know yet
This preprint has not been peer-reviewed. The researchers demonstrated the system on a single instrument in one oxidant-screening campaign. The efficiency results come from a simulation that replays archived data, not from new live experiments.
Terms in this story
- Self-driving laboratory
- An automated lab in which software and robotics plan, carry out and analyze experiments with little human involvement.
- Liquid chromatography-mass spectrometry
- A technique that separates the components of a mixture and then identifies them by their mass.
- DMTA cycle
- A repeating loop of designing, making, testing and analyzing that structures experimental work.
Sources and verification
DexScience articles are produced with AI assistance and an automated fact-check against the linked sources. Spotted an error? Tell us and we will correct it openly.


