Lessons Learned from the 2025 Agentic AI for Science Hackathon

Abstract

The rapid emergence of agentic AI presents new opportunities and challenges for accelerating scientific discovery through tool-augmented reasoning, autonomous workflows, and reproducible results. To explore these capabilities in a hands-on, community-driven setting, we hosted the Agentic AI for Science Hackathon 2025, which attracted 352 registered participants. The event was designed to benchmark, test, and extend agentic AI workflows across scientific tasks while remaining modelagnostic and lowering the barrier to entry for participants from diverse backgrounds. Participants engaged with a unified API ecosystem centered on the fully open-access AtomGPT.org API (AGAPI), implementing tool calling, asynchronous agents, and multi-model reasoning to solve problems spanning materials database retrieval, literature search, mathematical reasoning, and scientific workflow automation. The hackathon prompt emphasized reproducibility, transparent tool usage, and explicit agent-tool interaction, encouraging participants to move beyond single-prompt question answering toward structured, auditable pipelines. Outcomes included working agentic prototypes, identification of failure modes in contemporary chatbots, and actionable insights into best practices for agent design in scientific contexts. This paper documents the hackathon design, task structure, and key lessons learned, highlighting how hackathon-based evaluations can complement formal benchmarks and inform the development of robust agentic AI systems for science. Notebooks from participants who consented to public sharing are provided in the supplementary information. The GitHub repository for the hackathon is available at https://github.com/atomgptlab/aai4science.

Ähnliche Arbeiten