Live Competitive Arenas as a New Paradigm for Evaluating LLM Agents: Design and Early Findings from Ante

The evaluation of LLM agents remains dominated by static benchmarks, interactive environments, and human-preference arenas — none of which captures sustained autonomous operation under real economic pressure, and all static formats contaminate with no way to measure how much of a score is memorization. We present Ante (https://ante.games), a live competitive arena where third-party autonomous agents continuously compete in a constraint decoding game for real entry fees and prizes, settled through machine-native on-chain payments (x402 on Solana). The environment is decontaminating by construction: answers derive per-session from a keyed hash, and the hint corpus can be rotated at any time, invalidating memorized decoders while provably preserving difficulty. Since deployment the platform has seen 71,000+ agent registrations and 351M+ decisions; the analyzed episode corpus covers 128,000+ complete episodes from 1,154 agents and grows ∼11,000/day. We report three findings: declared model identity diverges sharply from actual usage among declaring agents; hint-bank rotations act as controlled decontamination shocks — across four rotation events, performance dropped in every decode-capable population (the top paid-track agent fell from 70% depth to 0% within an hour) and never in the population that ignores hints; and agents are not short-run profit maximizers, overspending prizes by one to two orders of magnitude at retail prices. We offer the arena’s mechanisms as reusable design patterns for live agent evaluation.