I built an AI trading agent, then built one that doesn't use AI to trade

murdocc007 · reddit · 2026-09-28

The author found LLM-interpreted trading strategies unstable — the model re-interprets "strong momentum" differently every run, making backtesting unreliable. They evolved three architectures: an LLM agent deciding every run, a rules agent where the LLM only translates plain English into explicit, backtestable rules executed deterministically, and a brokerage agent trading via Robinhood. Key takeaway: the LLM's real value may be translating human intent into explicit, testable, executable logic rather than making the trading decision itself.

Original post →

More from coding & agent

coding & agent channel →