Options Alpha Agent

Options Alpha Agent is an autonomous options-trading agent built for the Alpaca AI Trading Agents Hackathon. Most trading agents ask which way the market will move — nobody answers that reliably. This agent asks a measurable question instead: is this options structure priced better than the risk it carries? We enumerated 60 real SPY/QQQ/IWM spreads and scored expected value under the market's own implied volatility. Zero were positive-EV — what an efficient market looks like — so any edge had to be real: the variance risk premium. Implied vol persistently exceeds subsequently realised vol, since sellers are paid to bear variance risk. The agent prices with implied vol but scores probability with realised vol; the gap is the edge. Re-scored this way, 5 of 12 candidates cleared a 2%-of-risk threshold; QQQ (implied below realised) correctly yielded nothing. The LLM stays out of the money path. Only two of nine deterministic cycle stages touch a model: a directional view and a structural critic, via Featherless AI (zai-org/GLM-5.2). Regime classification, ~37-structure enumeration, N(d2) EV scoring, 22 risk gates, and execution are deterministic, tested Python. The model only shifts probability, never a strike or size — which also defends against prompt injection, since a bad response can only produce a wrong opinion the EV test must justify. Trading live surfaced defects a backtest missed, now fixed and covered by hundreds of tests. Risk is layered: structure gates, sizing caps, a drawdown circuit breaker, and a broker-side halt that cancels orders, flattens the book, and sets suspend_trade on the account — verified against a real 403 rejection. It trades exclusively options as four-leg mleg orders, runs unattended over Alpaca's REST API, and uses Alpaca's MCP server for research and oversight. Broker-agnostic in principle, Alpaca is the first integration. The real product is the rejection log: why the agent stood aside. Paper simulation; not investment advice.