Skip to Main Content

Cursoe-tron

Built at The Recursive Agents Hackathon · Oct 17, 2026 · San Francisco, CA

Demo video · drive.google.com/…

Cursoe-tron Stunt School is a team of AI agents on Crusoe Managed Inference that teaches itself to fly a jet-powered car through stunts in a real 3D game, and it use N+1 to learn from the task till it able to do the real task like C-loop. From a one-line goal, a spec writer (Nemotron) writes executable checks, and an independent auditor (NVIDIA Nemotron-3-Super) attacks them for loopholes. A pilot (DeepSeek-V4-Flash) writes the flight program, a verifier runs it in the game's own physics, and a watchdog catches loops and drift, then decides whether to continue, re-plan, revert, stop or ping a human. Between rounds a coach runs physics experiments, rewrites the team's playbook, and must prove every new skill with a test flight; the tests are locked, so it can't game them. A rewrite is kept only if it measurably helps. On 4 stunts the team never trained on, the hidden referee's pass count went from 0/19 flights to 5/20 (mean score +85%, p = 0.02), with zero false "done" claims. Every flight replays in the 3D game

Team