The Wall Street Journal has a piece by Hannah Erin Lang on ordinary US investors wiring their brokerage accounts to agents built with Claude or Codex, then letting the resulting code trade for them. No quant background, no firm, no risk desk. A prompt, a broker API key, and a strategy that exists because a model wrote it.
Retail algo trading is not new. Backtesting platforms and copy-trading services have been around for years, and they mostly failed to spread because writing the strategy was the hard part. That barrier is gone. The person now supplies an idea in plain English, and the model supplies the loop, the order logic, the position sizing, and something that looks like error handling.
Where this actually breaks
The generated code usually runs. That is the problem. A strategy that executes cleanly on paper can still be silently mispriced, overfitted to the sample it was described against, or missing the one guard that matters when a position moves against you at open. The model has no view on whether the idea was any good; it has a view on whether the syntax parses.
Broker APIs make the failure mode worse, because the distance between a bad loop and a filled order is now roughly zero. Anyone doing this should be honest that they are running unaudited software against a live account, with real money, and that the losses land on them alone. Nothing here is investment advice, and I would treat any strategy generated this way as a hypothesis rather than a system.
The part that will stick
Firms have spent a long time selling the idea that systematic trading is inaccessible. This trend punctures that, at least at the surface. What it does not puncture is the rest of the stack: data quality, execution costs, slippage, the discipline to turn a losing model off. Those were always the moat, and they still are.
Expect brokers to notice. Rate limits, kill switches, and API terms that specifically address autonomous agents are the obvious next move.