What they did
The author runs a small spot-only crypto trading bot as a side project and handed its daily operations to Claude. Every evening the agent checks the bot, runs analyses on past trades, writes a daily report and proposes changes. It may not change trading logic by itself. The author also gave it a short plain-language rulebook, including one rule that logic only changes once an edge is confirmed, and that a rejected idea must be written down with the reason. When the agent saw most losses came from round trips of a few hours, it backtested a minimum holding period on about four months of 15-minute candles, trying 17 variants.
How it went
The backtest looked like a win: no hold gave 178 trades and about -10 cumulative, an 8 hour hold about +9, a 24 hour hold about +22 with a 50% win rate. The agent then tried to break it. Bootstrapped 95% intervals crossed zero for every setting, the latest period lost money for all of them, simply holding the asset made about +44, and random entries did about as well as the current rules. Nothing was deployed. The agent instead started a forward-only comparison of the 24 hour variant, to be judged after 30 round trips or four weeks.
Worth knowing
The agent has no memory between evenings, so the author makes it log rejected ideas and the criteria for revisiting them in a Decisions and reasons section of each report, or the same tweak returns every day. The author credits the written rule, not the model, for stopping the change. The harness running Claude was not named, and the forward test had no result yet.