At our Claude Conversation, Berlin scientists named the open problems in their research they most want AI help with. In this Impact Lab we build on them.

The Impact Lab is a full-day build. Teams form around research problems that grew out of the Conversation. The day starts with a short introduction, examples of research loops, and a presentation of Claude Science. Then each team builds a research loop for its problem: an agent that reads the literature, writes and runs code, checks its own results, and keeps going on a problem a scientist actually cares about. Experienced mentors from the Claude Community support the teams throughout the day.

We have seen that a loop only starts to produce value when the scientist who owns the problem works directly with the engineer who builds the harness. This day puts the two side by side.

The series is co-hosted by the Foresight Institute Berlin AI Node. It continues Foresight's Existential Hope work on how advanced technology can go well for people, turning the problems scientists actually face into open tools anyone can keep using.

How it works

  • Before the day: apply, and we select participants and match teams to the research problems from the Conversation. Whoever brings a problem prepares a short brief and any public or anonymized data. No code is written beforehand.

  • On the day: a short introduction, examples of research loops, and a presentation of Claude Science, then 6–8 hours of building, with experienced Claude Community mentors supporting the teams.

  • Closing: every team demos its loop: what it found, where it got stuck, and what it would need next. A panel including the host, a Foresight Institute representative and, if possible, an Anthropic judge picks the winning team.

What you leave with

  • A working, open-source research loop on GitHub, aimed at a real open problem and free for the scientist and anyone else to keep using

  • $100 in Anthropic API credits per participant (requires a free account at console.anthropic.com; bring your Organization ID)

  • One month of Claude Max 20x for each member of the winning team

  • Collaborators across the line between science and AI engineering

Who it's for

  • Domain scientists with an open problem from their own work, whether or not they came to the Conversation. Minimal programming experience is enough. Problems fit best where results can be checked computationally, for example in protein and molecular design, chemistry and materials, mathematics, genomics, nanotechnology, physics, climate science, biomedical imaging or neuroscience.

  • AI and harness engineers comfortable with coding agents like Claude Code, who want to work on problems beyond software.

Bring: a laptop (8 GB RAM minimum, Python 3.10+), a console.anthropic.com account, and for scientists, your problem brief and data.

Free admission. Participants are selected by application. Hosted by Kazik Pogoda, AI researcher and volunteer Claude Ambassador for Science working with the Foresight Institute on epistemic AI tools for science, and Keith Patarroyo, Foresight Institute fellow at the University of Glasgow. Together they presented the AI4Science Buildathon concept that won a prize at Foresight's 2026 Secure & Sovereign AI Workshop, and this series is its first edition.