Karpathy showed why most LLM trading bots lose money and how to fix it with 7 rules
The problem is that the LLM decides everything on its own exits, leverage, which coins to trade, and loses money as a result. A lab put 3,505 agents on real money and most of them lost
The solution is called a guardrail layer the LLM only picks trades, while code handles a hard 2% stop and 4% take-profit, leverage that shrinks when the market gets crazier, and a second smaller LLM that checks every order before it fires
Seven rules: hard-coded exits, leverage computed in code, a tested coin list, deleting tools that change nothing, a checker LLM, logging what the model saw before each trade, and replaying trades on a cheap LLM before paying for a smarter one
Result 71 of 205 liquidations caught by flagging just the riskiest 10% of orders, and three frontier LLMs making equally good calls while one costs 25x less
The catch none of the agents found an edge these rules just keep you alive while you hunt for one