What they did
The author set up a $1M paper account at Interactive Brokers and split the work into roles on one small cloud server. Four AI agents each run on their own weekday: a head of research looks for ideas, a strategist turns an idea into a rule and backtests it, a validator tries only to disprove it, and a CIO reads the week and writes a Friday letter. Four control steps are plain code, not AI: operations reconciles against the broker, compliance checks the instrument is allowed, risk checks drawdown limits, and execution places the order only after the other three pass. The strategy is a simple trend rule with volatility sizing and weekly rebalancing, kept in a versioned repo with a shared agenda and lessons log. Every role sends notifications to the author's phone, which also carries a kill switch.
How it went
The first live order was rejected because a UK retail account may not buy many US-listed funds, so a compliance check was added and the universe moved to European funds. Later a test on 2022 to 2026 data showed the trend timing earning about half the risk-adjusted return of just holding. The validator agent pointed out that the sample had no crash, and a rerun covering 2008 and 2020 reversed the result: about +7.6% against -14.8% in 2008. The kill switch failed silently at first because of a config error. The author makes no profit claim; the book has run only a short time.
Worth knowing
This is paper money, not real capital. The agents do not learn between runs; improvement comes from the written lessons log every agent reads. The author does not name the model or agent framework used, and the hard limits (5% drawdown pauses orders, 7% goes to cash) live in code the agents cannot argue with.