What they did
Sayantan Sanyal built a paper trading simulation inside Claude Code. Three subagents each hold their own $50K fake portfolio and a narrow outlook: a technical trader that reads RSI, MACD and moving averages, a fundamental trader that looks only at valuation, debt and margins, and a retail investor that follows news and momentum. A manager agent takes his questions and runs turn-by-turn debates in which each trader tracks a buy, hold or sell stance and a conviction level, with a 5-minute timeout. Market data comes from yfinance. A deterministic Python script places every trade and does all the arithmetic, and a reviewer agent logs results and feeds lessons back to the traders.
How it went
Across 5 rounds over 3 days, every trader lost money in what he calls a bad week for the market. The technical trader stayed mostly in cash and tracked the S&P 500 closely; the fundamental trader kept averaging into losers and trailed the index by four points; the retail trader lost to its own flip-flopping. In one debate on three tech stocks to buy, all three agreed on NVIDIA and Google within seven turns but hit the timeout still split between Broadcom and Microsoft.
Worth knowing
It is fake money and a short run, and the author says plainly it is not a trading tool or advice. His main design rule: the model decides, but a plain script owns position sizing and portfolio edits. The code is public on GitHub and needs no paid API keys beyond Claude Code.