Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Everyday tasks

A human-agent team ran X-ray experiments you can audit

Written November 12, 2026

Three physical synchrotron experiments, run by a human-agent team, with a complete transcript behind them. The record shows who proposed each step, who corrected each error, and where the humans stayed in charge.

What actually ran

Closed-loop autonomous experimentation promises faster science, but building the control software for a physical instrument is a project in itself, so few labs get there. A preprint by Gunik Luthra and eight co-authors, submitted on November 6, 2026, takes a shorter path: coding-agent infrastructure, grounded in validated scientific methods, running synchrotron X-ray experiments on nanoparticle superlattices. A synchrotron beamline is real, shared, expensive equipment, not a simulator, and the experiments ran under objectives that humans set and kept revising as results came in.

Across three experiments, the system kept operating on the physical instrument. That sustained operation is the paper's first claim, and it is the part most agent demonstrations skip: an agent that books beam time has to keep working when the sample, the beam, and the plan all change mid-run.

What auditable means here

The team kept complete transcripts of the human-agent and agent-agent exchanges. Those transcripts are the audit trail. From them, the authors reconstruct decision provenance: which participant proposed a step, whose input shaped the decision that was finally taken, and who caught which error. They use the same record to describe how the cognitive work split between people and agents. For a reader, that is the difference between "the AI did an experiment" and an account you can pull apart decision by decision.

The division of labour, in numbers

Between 61% and 67% of the messages across the three experiments involved interpretation or judgment, rather than routine execution. The character of the work also shifted as the team settled in: methodological and scientific contributions rose from 46% of messages in the first, setup-focused experiment to between 56% and 61% in the later ones. Error correction was mostly a team sport — between 59% and 68% of corrections involved multiple participants — and about half of all decisions incorporated the input of another participant, human or agent.

Read plainly, the agents were not left alone with the beamline and the humans were not supervising from a distance. Each decision that mattered tended to pass through more than one mind, and the transcript is what lets the authors say that with numbers instead of impressions.

Strengths

Limits, and what the agents did not do

Related playbooks

Sources

Back to: Everyday playbooks hub · Also: Contrastive Discovery: make hypotheses compete · Scientist World: 1,000 executable research tasks · AlphaGenome 2 reads variants per cell type