Crossposted from the Lightcone Research website
From July 28–31, 2026, at the AI Futures Lab in Berkeley, our first Agentic AI for Science (AAI4Science) Developer Summit brought together 35 scientists and builders spanning astrophysics, neuroscience, medical imaging, chemistry and metabolomics, geophysics, biology, theoretical physics, statistics, and data science. Our goal was to lay the foundations for an open-source ecosystem of AI-powered tools that support rigorous scientific research.
Over four days, participants took part in dedicated discussion sessions (summarized below) examining the key advances needed before agentic AI can be adopted broadly across science. They also began work on a dozen collaborative projects aimed at tackling some of these challenges.
As a major by-product of the week, a lot of thinking has gone into refining the Agentic Schema for Transparent Research Analysis (ASTRA) schema, with hack projects and discussions converted into Requests for Comments (RFCs) that will inform the ASTRA spec as we work toward a v1 release (stay tuned!).
We were amazed by the breadth of what participants built, from tools that trace every twist and turn of an experiment to an entirely new graph-based interface for exploring scientific papers. Visit the Lightcone Research blog for more details of the unconference discussions and team projects.
Explore the Unconference sessions
- What is the optimal way to share scientific knowledge in the age of AI and how do we get there?
- How to build long-running agentic sessions that perform research on HPCs following best practices?
- How to iterate with scientists on building a joyful experience?
- How do we democratize the access to frontier AI resources to our scientific community?
- How to verify the validity and quality of AI agents’ findings?
- Provenance from all angles: addressing the needs of both science and HPC
Learn more about the Projects
- Agentic Interaction Pattern (Steffen Bollmann, Eli Bressert, Jarrod Millman, Liam Parker, Fernando Pérez)
- Provenance-preserving composite figures (Pete Bachant, Nokome Bentley)
- ASTRA for multiverse analysis (Piyush Acharya, Steven Dillmann, François Lanusse, Anthony Ozerov, Jacopo Teneggi)
- Actor attribution for ASTRA decisions (Martin Legrand, Oliver Muellerklein, Benjamin Navet)
- Audience-aware skills (Ioana Ciucă, Cail Daley, Silas Santini)
- A personalized, shareable research notebook (Nolan Koblischke, Anthony Kremin, Huiqi Wang)
- Towards a verifiable agentic scientific HPC-as-a-Service (HAPS) architecture (Debbie Bard, Alexandre Boucaud, LogicA team, Corentin Val)
- Benchmarking agentic skills on HPC (Debbie Bard, Alexandre Boucaud, Kangning Diao, Louis-Félix Nothias, Corentin Val)
- Normalizing agent trajectories into ASTRA (Shruti Badhwar)
Get full AAI4Science summit details, including links to the slides!
BIDS is proud to support Lightcone Research, an international initiative building infrastructure for rigorous AI-assisted computational science.