Tufa Labs took first place in Milestone 1 of the ARC-AGI-3 competition on Kaggle!

ARC-AGI-3 challenges agents to solve unfamiliar interactive puzzles: they must discover the rules, infer the goal, and adapt as they play. Our winning solution, the Duck Harness, gives a language model a Python environment to inspect the game, test ideas, and take actions.

We’ve open-sourced the harness and shared a technical write-up covering its design, results, and what we’ve learned along the way. We’re excited to build on this first milestone and see what the community develops with it.

Read Duck Harness: Winning Solution for ARC-AGI-3 Milestone 1 for the full story, or explore the code on GitHub.