杰克·Nesler（Jake Nesler）的Claude驱动交易代理在最初做对了一件重要决定：11月末Nvidia财报使股价大幅上扬时，它选择了不追高。若追涨，此举据估算会在当周给其投资组合增加约1万美元（USD 10,000）的损失。29岁的Nesler在因日常工作之外交易期权后精疲力尽，受到Anthropic实验启发，花了约2.5周教会模型其风险偏好、入场信号和仓位规模规则，并在Alpaca模拟账户上用10万美元（USD 100,000）资金测试。实际表现起伏很大：五个交易日内仅有一个成功信号，随后接连出现投机性亏损。
尽管舆论热度很高，零售交易者正在快速采用AI代理。像OpenClaw这样的开源平台让用户可通过WhatsApp与Telegram与代理交互；在X上出现了大量夸张收益叙事，其中一条被浏览470万次的帖子声称Polymarket两天获利5860%，但随后被另一账号否认为不现实。该热潮在于交易者希望由AI替代手动流程，尤其是在Robinhood一类应用养成的交易人群中，而交易所层面也在顺应：Polymarket、OKX、Bybit与Kraken均已提供AI代理下单接口。
然而可持续盈利证据并不牢靠。由于模型来自大量金融建议与风险管理文本训练，它倾向于保守，默认更愿意买入蓝筹和标普500成分股，除非用户持续要求提高风险偏好。即便Nesler对其进行了调教，30天内仍仅报告约7%收益，略高于同期标普500约4.5%，但回撤最高达22%。他后来在Kalshi上投入约30美元测试体育事件，表现“像老虎机”一样：尽管在比特币价格区间交易中命中率约60%，终究最终亏损殆尽。研究者指出，交易优势在被大众共享后会迅速消散，若过多代理涌入预测市场，可能只是复读互相可见的信息，把市场变成共识回音室。
Jake Nesler’s Claude-driven trading agent got one big decision right early on: when Nvidia surged after earnings in late November, it chose not to chase the rally. Doing so would have added about USD 10,000 to his portfolio loss for the week. At 29, a burnout-prone options trader, Nesler spent about 2.5 weeks teaching the model his own rules for risk, entries, and position sizing, then tested it on Alpaca with USD 100,000 in simulated capital. The first week was volatile: one useful signal over five trading days, then repeated speculative losses.
The retail ecosystem has expanded fast regardless. Open-source systems such as OpenClaw now let users operate agents through WhatsApp and Telegram, while X has become crowded with return claims, including a post viewed 4.7 million times alleging 5,860% profit in two days on Polymarket, later challenged as implausible. Traders are attracted because AI automates workflows that once depended on people using apps like Robinhood. Exchanges also see upside: Polymarket, OKX, Bybit, and Kraken have all introduced interfaces for AI agents.
Evidence of durable edge remains weak. Models inherit conservative defaults from financial-advice training, often preferring blue chips and S&P 500 names unless users force more risk. Even after tuning, Nesler reported about 7% return over 30 days versus roughly 4.5% for the S&P 500, while drawdowns reached 22%. His later Kalshi test was near a slot-machine pattern: although Bitcoin bracket trades won about 60% of the time, it eventually lost all its capital. Quant experts caution that trading edges decay when shared, and if too many bots dominate prediction markets, they may only amplify consensus and convert forecasting into an echo chamber.