ARC Prize 2026: Adding AI capabilities lowered agent's competition score

A developer competing in the ARC Prize 2026 challenge — where agents must learn unseen games without instructions — found that expanding their agent's capabilities repeatedly caused its benchmark score to drop. After adding a fourth game-solving skill that was verified to work in testing, the agent's score fell to 0.04, its worst result yet, down from an already modest 0.09. A survey of 25 practice games revealed the core problem: the agent was spending moves probing games it ultimately could not solve, turning each failed attempt into a net loss. The issue was not the new skills themselves but the increased cost of attempting more unwinnable games. The developer concluded they had been optimizing for raw capability rather than net outcome, a distinction the competition's scoring makes unforgiving.