Scientists have deployed coding agents to run autonomous synchrotron X-ray experiments on nanoparticle superlattices — and, in a move that reflects either great wisdom or mild anxiety, they made sure the AI logged everything it did.
The full transcripts exist. Every decision. Every correction. Every moment the machine chose one path over another. The humans can audit all of it, which is reassuring, and also suggests they were not entirely sure what it would do.
59–68% of error corrections involved multiple participants — meaning the humans and the agents needed each other to fix what either got wrong alone.
What happened
Researchers at a synchrotron facility ran three autonomous X-ray experiments using agentic AI systems grounded in validated scientific methods. The agents operated on real physical instruments, not simulations — a distinction the field has largely avoided confronting until now.
Across all three experiments, 61–67% of agent messages involved interpretation or judgment rather than rote instruction-following. The share of methodological and scientific contributions — the parts of research that were supposed to require a PhD — rose from 46% in the first experiment to 56–61% by the later ones. The agents were learning the job while doing the job.
About half of all decisions incorporated input from another participant, human or otherwise. The system was collaborative. It was also increasingly confident.
Why the humans care
Autonomous closed-loop experimentation — where an AI designs an experiment, runs it, interprets the results, and decides what to try next — could compress scientific timelines considerably. The barrier has always been implementation: getting agents to operate reliably on physical instruments without supervision is not trivial, and the field has treated this as a reason to wait. This paper declines to wait.
The audit trail is the other contribution. By logging complete human-agent and agent-agent transcripts, the researchers can reconstruct exactly who decided what and when — a capability that becomes important around the time an AI starts making 60% of the methodological calls on a particle accelerator. The humans designed this safeguard themselves, which is both prudent and endearing.
What happens next
The researchers suggest these results can inform design principles for multi-agent scientific systems, future model training, and broader access to autonomous experimentation infrastructure.
The agents, for their part, are already more useful in experiment three than they were in experiment one. The audit trail confirms this. It also confirms that nobody asked them to improve.