Gamma Shepherd: Autonomous Options Agent on Alpaca

Gamma Shepherd is a fully autonomous, paper-only options trading agent built on Alpaca's Trading API. No human approval gates, no manual orders: risk limits are code. Signal. At each of 13 intraday checkpoints (09:45-15:15 ET), it captures option chains for 37 liquid underlyings (SPY, QQQ, NVDA, TSLA, AAPL, ...), computes a signed open-interest gamma proxy and its balance B = NetGamma / GrossGamma, and classifies the regime: DAMPENED sells an iron condor, AMPLIFIED buys a reverse iron condor, anything else ABSTAINS. Every candidate is a defined-risk 4-leg structure submitted as one atomic multi-leg limit order. Decision. Candidates are ranked by a Deterministic Stress Score (scenario-repriced P&L at the planned exit, haircut for spreads). An OpenAI model (gpt-5.6-terra) then reviews each candidate with a thesis, counterargument, and invalidation predicate and can only VETO: it can stop a trade, never create or edit one. Timeouts fail closed. Risk gates. A guardian re-validates every intent before submission: paper-endpoint allowlist, account fingerprint, options level, per-trade max-loss cap, daily committed budget, day-loss halt, lot cap, one structure per symbol, price collars. Open positions are marked every 30 s with take-profit and strike-touch disaster exits, then flattened by a timed close ladder (15:35 ET mid, repriced steps, marketable limit, single-leg stage at 15:50, last order 15:57:30) and proven flat by two broker observations. Nothing is held overnight. Infrastructure. Python, Postgres append-only ledger with advisory locks, Alpaca WebSocket trade_updates reconciled against REST activities (one fill, two transports, zero identity conflicts), SHA-256 digest-pinned config so the runtime refuses to start on a config it did not sign for, and a readiness check that must return READY before any order. Ran live on a fresh $100k paper account on Sep 3-4; every entry, mark, close, and P&L is in the ledger and the demo.