Auditable Human-Agent Teams for Autonomous X-ray Experimentation

Authors: Gunik Luthra, Ronald J. Pandolfi, Benjamin T. Diroll, Ricky Ronquillo, Ruipeng Li, Christopher B. Murray, Marcus M. Noack, Kevin G. Yager, Esther H. R. Tsai

Subjects: Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)

Cite as: arXiv:2611.12933 [cs.AI] · https://doi.org/10.48550/arXiv.2611.12933

Submitted: 6 November 2026

Abstract

Closed-loop autonomous experimentation can accelerate scientific discovery, but the implementation effort required limits its accessibility. Agentic systems may lower this barrier, although sustained operation on physical instruments remains largely unexplored. Here, using coding-agent infrastructure grounded in validated scientific methods, we demonstrate autonomous synchrotron X-ray experiments on nanoparticle superlattices under evolving, human-defined objectives. Complete human-agent and agent-agent transcripts enable an audit that reconstructs decision provenance and characterizes the division of cognitive labor.

Across three experiments, 61–67% of messages involved interpretation or judgment, while the share of methodological and scientific contributions increased from 46% in the setup-focused first experiment to 56–61% in later experiments. Provenance analysis showed that 59–68% of error corrections involved multiple participants and that about half of all decisions incorporated the input of another participant. These results may guide design principles for multi-agent scientific systems and inform future agent harnesses, model training, and accessible autonomous experimentation.

Key Takeaways

• Lowering the barrier to autonomous science: Coding-agent infrastructure grounded in validated scientific methods enabled autonomous synchrotron X-ray experiments on nanoparticle superlattices, operated under evolving, human-defined objectives.
• Auditability by design: Complete human-agent and agent-agent transcripts allowed researchers to reconstruct decision provenance and characterize how cognitive labor was divided across the team.
• Judgment-intensive workflows: In 61–67% of messages across three experiments, participants engaged in interpretation or judgment rather than purely mechanical tasks.
• Growing scientific contribution over time: The share of methodological and scientific contributions rose from 46% in the setup-focused first experiment to 56–61% in later experiments.
• Collaborative error correction: Provenance analysis revealed that 59–68% of error corrections involved multiple participants, and roughly half of all decisions incorporated input from another participant.
• Design implications: The findings offer guidance for multi-agent scientific system design and inform future agent harnesses, model training, and accessible autonomous experimentation.

Significance in 2026

As agentic AI systems move from software-only environments into physical laboratories—part of a broader push toward self-driving labs and AI-driven scientific discovery—questions of trust, traceability, and human oversight become central. This work demonstrates sustained, auditable operation of agent teams on a live synchrotron beamline rather than short proof-of-concept runs. By quantifying how cognitive labor is distributed between humans and agents—and how often decisions draw on multiple contributors—the study provides an empirical foundation for designing the next generation of multi-agent scientific systems, where auditability is treated as a first-class requirement rather than an afterthought.

via ArXiv AI