I made my agent more capable and it got worse

Short summary

A builder journal from the ARC Prize 2026 competition reveals a counterintuitive lesson: adding more capabilities to an AI agent actually lowered its score. Each new skill caused the agent to attempt more games it couldn't solve, wasting moves on failed theory-building. The key insight is that net result matters more than raw capability — a bigger toolbox is only an asset if you know when to keep it shut.

• •Adding a fourth game-solving skill dropped the agent's score from 0.09 to 0.04
• •The cost of attempting unsolvable games outweighed the benefit of new capabilities
• •Core lesson: optimizing for capability instead of net result is a trap — restraint, not power, drives better outcomes

Generated with AI, which can make mistakes.