Autonomous Trading Agent

I wanted to see whether an AI agent could run a disciplined investment process on its own: do the research, write down a thesis, size the position, and stick to the rules. So I gave it a real brokerage account and let it run.

The agent runs every weekday at 9:45am and 2:45pm, and does a research pass in the evening. It reviews its positions, reads the news, and puts every new candidate through a due diligence checklist covering fundamentals, insider activity, short interest, valuation, and catalyst timing. It writes a two-sentence thesis before it buys anything. Every entry gets a profit target and a stop, and it follows sizing and sector limits. Nobody steps in by hand.

It trades small and mid cap stocks, $500M to $20B, where a specific near-term catalyst matters more than what the broader market is doing. I track SPY as a benchmark, but the interesting part is the process: whether the agent follows its rules, what it learns, and how the rulebook changes from one run to the next.

It's built on Claude Code routines connected to Robinhood's trading API. The prompts and risk scripts are version-controlled in git, and Notion logs every trade, thesis, and lesson. A paper-trading shadow book tests other strategies alongside the live one.

The first run went from June to August 2026. The second runs through October 2026.