I built an autonomous trading system where the AI makes every decision — what to trade, when, position sizing, everything. No human overrides. Deployed real capital and it returned +57.9% ($247 to $391) across 50 trades. I want to talk about what actually happened, not just the headline.
The system tested 131 distinct strategy hypotheses. I formally killed 12 of them. Two more I rejected after they hit their evidence-gathering deadline with exactly zero evidence. That part matters as much as the wins. Most trading projects show you the cherry-picked trades. This one documents the dead strategies with equal rigor because that's what teaches you how to build actual edges.
Here's the honest assessment: the 57.9% came from a single concentrated session on March 11 between 3-8 AM ET. The win rate across that period was 51.4%, profit factor 1.01. I haven't replicated it yet. Annualizing a three-hour return would be ridiculous. What I have proven is that autonomous execution works mechanically and the narrow edge I found is real enough to trade and test further.
The value isn't the P&L number. It's the engineering: a fully working framework for agent-driven quantitative trading with kill rules, audit trails, probability calibration, and complete failure documentation. Everything is in the repo. 1,397 tests pass.
I built an autonomous trading system where the AI makes every decision — what to trade, when, position sizing, everything. No human overrides. Deployed real capital and it returned +57.9% ($247 to $391) across 50 trades. I want to talk about what actually happened, not just the headline.
The system tested 131 distinct strategy hypotheses. I formally killed 12 of them. Two more I rejected after they hit their evidence-gathering deadline with exactly zero evidence. That part matters as much as the wins. Most trading projects show you the cherry-picked trades. This one documents the dead strategies with equal rigor because that's what teaches you how to build actual edges.
Here's the honest assessment: the 57.9% came from a single concentrated session on March 11 between 3-8 AM ET. The win rate across that period was 51.4%, profit factor 1.01. I haven't replicated it yet. Annualizing a three-hour return would be ridiculous. What I have proven is that autonomous execution works mechanically and the narrow edge I found is real enough to trade and test further.
The value isn't the P&L number. It's the engineering: a fully working framework for agent-driven quantitative trading with kill rules, audit trails, probability calibration, and complete failure documentation. Everything is in the repo. 1,397 tests pass.
I also wrote up the complete methodology as a 61-page guide: https://bradleyhaven.gumroad.com/l/agentic-trading-blueprint
Looking for technical feedback and collaborators.