ARC-AGI-N: open-source tool finds open math and science problems for AI agents

Physical_Pepper6294 · reddit · 2026-09-08

Inspired by a non-mathematician's Claude attempt at the Riemann hypothesis (which still improved the lower bound on zeta zeros on the critical line from 41.6% to 67.2%), an 11-day Lean formalization of Wiles's FLT proof, and "GPT-Astra scoring 99.9% on ARC-AGI-3", the author built ARC-AGI-N — a research tool that finds open problems in math and science and prepares context for AI agents to take a stab.

Features:

Stack: Next.js/React/TypeScript, OpenAI Luna + Valyu search and DeepResearch APIs. Open-source, self-hostable, hosted version available. The author is upfront that it doesn't turn prompts into valid proofs — the point is surfacing worthwhile attempts with enough context to start. Researchers with neglected problems are invited to contribute.

Related event: ARC-AGI-N: Open Tool Auto-Discovers Open Problems for AI Agents(2 posts)→

Original post →

More from coding & agent

coding & agent channel →