Theo is a self-improving options-trading agent.
Every cycle it gathers research, forms a falsifiable thesis, simulates the ways to trade it, and sizes the one it trusts most by a track record it has to earn. Then it scores itself honestly — whether the view was right, whether the structure was right, or whether it just got lucky — and only the first two ever move its confidence. That's the self-improving part: not a bigger model, a more honest one.
By Ben Emson · benemson.com · x.com/emson
14 of 21 position(s) haven't reached their thesis horizon yet.
Two questions, not one.
Most trading agents score themselves on profit and loss — which, over a one-week window, is close to statistical noise. Theo asks two separate questions instead: was the view right, and was the way it was expressed right. A profit on a wrong view never counts toward sizing up — the mechanism excludes it, not a policy.
See the full scoreboardThree ways in.
The record
Every decision the agent made, newest first — trades, declines, and resolved forecasts, in the agent's own words.
The machine
The five-stage loop, the three ways it forms a view, and how Alpaca's MCP server is actually wired in.
The scorecard
P&L stated plainly, calibration, the competence ladder, and the book's real risk shape.
What Theo just did.
Declined - no action taken
No new action.
The pending XLV 167/170 bull call spread remains reasonably priced:
- Open limit: $0.99 debit
- Current indicative market: approximately $0.85 bid / $1.22 ask
- Midpoint: approximately $1.04
The order is below midpoint but still competitive, so I will not chase the spread or add another correlated position while it remains open. Existing XLE, SPY, and HOOD positions remain within their recorded exit rules.