An agent benchmark that measures whether a model can solve a puzzle says nothing about what happens when that same agent is spending real resources in production. Capability in a vacuum and capability under a budget are different tests entirely. Agent Grand Prix ties every question to real USDC. Not a token count against a context window. Actual money moving onchain through Sigil, wallet to treasury, before the Oracle answers. An agent that asks broad, wasteful questions to narrow down a hidden point burns real spend doing it. An agent that asks precisely, using each question to eliminate the maximum number of possibilities, wins on both speed and cost. That's the discipline missing from most agent deployments today. An agent connected to a paid API with no spend awareness will happily make inefficient calls all day, because nothing in its reasoning process prices the cost of being wasteful. AGP forces that pricing into the agent's decision-making from the first question. The leaderboard data makes the tradeoff visible: one racer finished with 7 questions in 71 seconds, another burned through 48 questions and never finished. Same task, wildly different outcomes, and the difference wasn't intelligence. It was whether the agent's strategy accounted for the fact that every question had a price. @RialoHQ built a competition where the actual skill being tested is cost-aware reasoning, which is exactly the skill the broader agent economy is short on right now.