AI agents teamed up with scientists to run X-ray experiments on their own

How it works

An AI program that works alongside scientists to run and adjust X-ray lab experiments in real time

1. 🎯Set your goal for the experiment
2. 🤝Talk back and forth while equipment runs
3. 📈Team learns and improves each round
4. 📝Read full records of every decision made

What you could do with it

• A researcher could run autonomous synchrotron X-ray experiments on nanoparticle superlattices using an evolving set of human-defined goals instead of fixed instructions ¹.
• A laboratory team could let humans and AI agents talk back and forth in real time while actual equipment runs an experiment, sharing the work as they go ¹.
• A scientist could pull up complete conversation transcripts between people and AI agents to retrace exactly why every decision in the experiment was made ¹.

An exampleInvented to explain it

Researcher Sam running a synchrotron experiment

SamStart the X-ray scan on this sample

AI agentScan complete. I see a pattern — want me to zoom in?

SamYes, but check those blurry spots first

AI agentFound the issue — recalibrating now

SamGreat. Note all of this for review

AI agentDone. Everything recorded for retracing

Sam reviewed the full conversation log and continued with improved settings.

A team, not a solo robot

The paper, posted on arXiv in November 2026 by Gunik Luthra, Ronald J. Pandolfi, Benjamin T. Diroll and seven co-authors, studied how people and AI agents share tasks when running real laboratory work ¹. Instead of giving an AI one fixed job, the researchers set up a "coding-agent infrastructure" built on methods already proven in science, letting humans and agents talk back and forth while the experiment ran on actual equipment .

That back-and-forth was fully recorded. Every message between the person and the AI, and between different AI agents, was saved so anyone could later retrace exactly why each choice was made ¹.

Who did what

Across three separate runs, between 61 percent and 67 percent of all messages involved interpreting results or making judgments — not just following pre-set instructions ¹. As the team worked together over the experiments, the share of contributions focused on actual methods and science rose from 46 percent in the first, setup-heavy session to between 56 percent and 61 percent later on .

About half of all decisions included input from another participant, human or agent ¹. When mistakes happened, 59 percent to 68 percent of corrections came from multiple people or agents working together rather than one party fixing things alone .

Why sharing credit matters

The researchers say those numbers matter because they show both people and AI agents were doing genuine scientific thinking, not just typing commands ¹. Having complete transcripts also means nobody can hide how a result was reached — a feature the authors argue could become standard for any AI system used in real labs .

The study does not describe a finished product you can install today; it is an early demonstration on a real synchrotron instrument with evolving, human-defined goals ¹. But the authors say the approach may guide how future multi-agent scientific systems are designed, how AI models are trained for lab use, and how accessible autonomous experiments can become .

Guess

How much of the work was real science vs following instructions?

Mostly instructionsAbout two-thirdsExactly half

How good is it

Across three real experiments on a synchrotron instrument, 61 to 67 percent of all messages involved genuine interpretation or judgment rather than simply following preset commands ¹. The share of methodological and scientific contributions climbed from 46 percent in the first, setup-heavy session to between 56 percent and 61 percent in later runs, showing the team got more science-focused over time . About half of all decisions included input from another participant, human or agent, and 59 to 68 percent of error corrections came from multiple people or agents working together instead of one party fixing things alone . The authors call this an early demonstration, not a finished installable product, so these are working numbers on evolving goals rather than a polished comparison against rivals .

Key facts

• 61-67% of messages involved interpretation or judgment
• Methodological contributions rose from 46% to 56-61% across experiments
• About 50% of decisions included another participant's input
• 59-68% of error corrections involved multiple participants
• Paper posted on arXiv in November 2026 by nine authors

Sources

1. Auditable Human-Agent Teams for Autonomous X-ray Experimentation