Embedded Files
Designing agentic AI for how science gets done
Designing agentic AI for how science gets done
Researchers, infrastructure managers, developers, data curators, and funders all need different things from agentic AI. But scientific work is not accomplished alone. Systems must support dynamic teamwork and varied forms of collaboration—not just one person working with one agent. SciData’s AI UX team is studying those needs as agentic AI evolves and takes hold in scientific work. Our mission is to systematically research evolving scientific practices, translate those insights into core design principles, and leverage both to craft tools that help others build better AI systems for science.
Research Efforts
Research Efforts
AI user needs in modern data facilities
AI user needs in modern data facilities
What would make AI useful across the work of a modern data facility? We interviewed 18 researchers, developers, and other stakeholders about how they use AI today.
Some lessons learned:
Connect context, but give people control. Scientific data, project records, and reusable AI skills can make agents more useful. Who gets access—and what gets shared with teammates or agents—needs careful design.
Speed up the work, not past the thinking. Participants valued moving faster, but worried that rapid results could encourage weak ideas or erode skills. Build in opportunities to pause, examine, and question the work where they matter.
Access shapes what’s possible. AI use is often limited by cost or the availability of institution-provided tokens, not just by whether a tool is useful.
Learning from the BERIL Co-Scientist
Learning from the BERIL Co-Scientist
At a one-day workshop in May 2026, 38 researchers used BERIL to explore their own research questions. Follow-up interviews with 11 participants revealed what it takes for an AI co-scientist to fit into real scientific work.
Some lessons learned:
Let researchers inspect data before using it. People need to understand what a dataset contains—and whether it can answer their question—before an agent builds an analysis around it.
Make room to change course. Research is iterative. Agents should show their progress and offer meaningful checkpoints where people can review data, scripts, and results.
Make the important moments stand out. Repetitive permission requests can bury consequential decisions. Clear signals help researchers focus their attention where it matters.
Investigating team use of AI
Investigating team use of AI
Our ongoing interview study is investigating the collaborative work of domain scientists and their use of agentic AI. Learn more here.
Learn more
Learn more
Results from our research are available in following tech reports and presentations:
Cohoon, J., et al. (2026). Toward Calibrated Trust: What Working Scientists Need from an AI Co-Scientist. Zenodo. https://doi.org/10.5281/zenodo.20595967
Cohoon, H., & Paine, D. (2026). Understanding the State of Scientific AI: User Needs and Design Implications. UC Berkeley. Report #: LBNL-2001613. https://escholarship.org/uc/item/8d32b7dw
Cohoon, J., & Paine, D. (2026, October 15). Beyond Editing Code: Uses for AI Tools in Scientific Work. Zenodo. US Research Software Engineering Conference 2026 (USRSE'26), San Jose, CA. https://doi.org/10.5281/zenodo.23400189
Acknowledgements
Acknowledgements
Content on this page was drafted with support from AI tools.
The Advancing User Experience Impact for AI-enabled Team Science Ecosystems (Democratizing Collaborative User Experience) project is supported by the Laboratory Directed Research and Development Program of Lawrence Berkeley National Laboratory under U.S. Department of Energy Contact No. DE-AC02-05CH11231.
The High Performance Data Facility (HPDF) is supported by the U.S, Department of Energy, Office of Science, Office of Advanced Scientific Computing Research under contracts 89243126CSC000213 (TJNAF) and DE-AC02-05CH11231 (LBNL).
The Biological and Environmental Research Intelligence Layer (BERIL) project is funded by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research under Award Numbers DE-AC02-05CH11231, DE-AC02-06CH11357, and DE-AC05-00OR22725.
The DOE Systems Biology Knowledgebase (KBase) is funded by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research under Award Numbers DE-AC02-05CH11231, DE-AC02-06CH11357, DE-AC05-00OR22725, and DE-SC0012704.
Page updated
Report abuse