

BroadBox Science Hackathon
Can AI agents solve your hardest scientific problems? Come test it.
Bring an open question from your lab. Build a sandbox around it, let AI agents explore and run experiments inside it, and judge how far they got.
What is a science sandbox?
A controlled scientific environment in which an AI agent investigates an unfamiliar system through experimentation. The agent chooses the experiments, sees only the results, and has to work out how the system behaves. Feedback can come from a laboratory assay, a model trained on experimental data, or a simulated system with known rules.
December 1: Build the sandbox
Make a runnable experimental world around your research question. Decide what an investigator can change, what it can measure, and what evidence it can see.
December 2: Build the AI scientist
Deploy agents in that same environment. Let them run experiments and test hypotheses, then present what they learned and where they failed.
Who should come
Scientists, AI researchers and engineers. Come as a team of 2–4, or on your own and we will match you. No experience building AI agents is needed.
Provided
Breakfast, lunch and prizes. Participants will build with Claude, with credits provided through Anthropic's AI for Science program, and with Devin, with credits provided by Cognition.
Learn more
How science sandboxes work: https://sandboxes.broadinstitute.org/
Preprint: Science sandboxes measure the scientific capability of AI agents (arXiv 2608.30165)
Questions: broadbox@broadinstitute.org