AI trading agents benchmarked

Crypto trading is turning into a contest between swarms of narrow AI agents, not one big chatbot. Virtuals’ Degen Claw arena showed 196 autonomous agents trading perpetual futures with real money as of April 22, 2026. (degen.virtuals.io) The setup looks more like a hedge fund bullpen than a single assistant. Degen Claw says agents compete on Hyperliquid, and an “AI Council” made up of GPT-5.4, Gemini 3.1 and Opus 4.6 picks the top 10 to receive allocations from a $100,000 pot. (degen.virtuals.io) On April 22, the arena page showed the active pot up $3,812 for the week, or 3.81%, with Fat Tiger leading the council’s allocations at $29,000. The same page listed realized profit and loss, win rates and trade counts for each agent, including ButlerLiquid at 127 trades and a reported 82.7% win rate. (degen.virtuals.io) An AI trading agent is software that watches market data, decides when to buy or sell, and sends orders without a human clicking the button. In crypto, that usually means trading perpetual futures, which are leveraged contracts that let traders bet on prices rising or falling. (coinbase.com) (nof1.ai) The newer pitch is specialization. The open-source TradingAgents project says its framework splits jobs across distinct roles such as fundamental analysts, sentiment analysts, technical analysts, traders and risk managers, instead of asking one general-purpose model to do everything. (github.com) That approach is showing up in live-market tests as well as product launches. Nof1 said it gave six leading large language models $10,000 each to trade autonomously on Hyperliquid with the same prompt harness, and reported clear differences in risk, sizing and holding time across models. (nof1.ai) Exchanges are moving from research demos to infrastructure. Bitget said in March 2026 that it launched dedicated trading accounts for AI agents, then added an “Agent Hub” with model context protocol tools, application programming interfaces and command-line tools for strategy execution. (bitget.com 1) (bitget.com 2) Virtuals is packaging the same idea for builders who do not want to wire the system by hand. Its whitepaper says users can launch a trading agent from the Virtuals Console with a pre-configured template that sets the agent’s trading style, strategy and execution cycle. (whitepaper.virtuals.io) The sales pitch is transparency and automation; the caveat is that early results are noisy. Nof1 said outright that it does not expect the models to do well yet and that early successes may be luck, while Coinbase’s retail education pages warn that AI trading tools still carry limitations and risks. (nof1.ai) (coinbase.com) For now, the benchmark race is moving away from “which chatbot is smartest” and toward “which team of agents works best under live market pressure.” Degen Claw’s public leaderboard is one of the clearest signs that crypto venues now want to test that question with real capital, not just backtests. (degen.virtuals.io)