OWL Agent — trades on rails, AI on a leash

Retail options income should be boring: sell cash-secured puts, take profit at 50%, roll for credit, sell covered calls if assigned. In practice it is manual, emotional, and error-prone — and the usual "AI trading agent" (an LLM picking trades) makes it worse. LLMs hallucinate, drift, and cannot be held to a risk mandate. OWL Agent inverts that design. A deterministic engine does 100% of the trading on hard-coded rails — the classic wheel — live and unattended on Alpaca paper trading since day 1 of the hackathon. Above it sits a Claude layer with exactly two powers: read the market regime, and veto a new entry. It can never place, size, or force a trade, and it can never block an exit. An AI that can only say NO is an AI you can actually trust on a broker. The audience is retail traders earning options income. Evidence: a 2.5-year backtest on real Alpaca option-trade bars (live-engine rules returned +32.2% with drawdown control; a pre-committed selection-first variant beat SPY +54.6% vs +45.6%), a read-only hosted dashboard with a live stress test, and a public decision journal — every order, every regime read, failures included.