When Agents Trade: Live Multi-Market Trading Benchmark for LLM Agents (AMA)

AMA ran four agent architectures — InvestorAgent from InvestorBench, TradeAgent from TradingAgents, HedgeFundAgent and DeepFundAgent — across GPT-4o, GPT-4.1, Claude-3.5-haiku, Claude-sonnet-4 and Gemini-2.0-flash, taking one decision per agent per day on TSLA, BMRN, BTC and ETH from 1 August to 30 September 2025, with a majority-vote ensemble alongside them. Market context is aggregated from web search, Finnhub, NewsData, yfinance, CryptoNews, Binance, Twitter and Reddit, and two domain experts verified the news summaries over 20 sampled days of BTC and TSLA coverage. It comes from The Fin AI, the group behind PIXIU/FLARE and FinBen.

The headline figures need their baseline beside them: the paper reports a Sharpe of 6.47 for InvestorAgent on TSLA and, for plain buy-and-hold over the identical window, a 46.88% return at a Sharpe of 6.00 on the same ticker. This is a two-month self-reported live window on four assets, so read “lifelong” and “real-time” as design intent rather than demonstrated longevity.

The released testbed orchestrates but does not contain the agents: each must be wrapped as an HTTP endpoint, registered in the testbed config and driven over a POST interface, and a run needs seven API keys. The repository is four scripts plus configs, with a single commit from January 2026 and no LICENSE file — its README offers only “provided as-is for research and evaluation purposes”, which is not an open-source grant, so the code is not legally reusable. The preprint itself is CC BY 4.0, and the separate Hugging Face leaderboard Space declares Apache-2.0. A WWW 2026 publication is claimed for the paper, but its DOI does not resolve and the record is absent from Crossref as of August 2026, so it is listed here as an arXiv preprint. The leaderboard Space reports itself as running, though the paper’s cited live URL did not respond when checked.