STONKS is an autonomous multi-agent AI options desk for Alpaca's paper environment. Eight specialized agents run an unattended 30-minute cycle: deterministic code analysts (trend, IV rank, gamma-weighted dealer exposure, liquidity, event risk) and a sentiment analyst that reads real Alpaca news with per-source credibility weighting and citations; a bull/bear debate (Toro vs Ursa, 2 rounds, claims must cite analyst facts); a judge (Verdi) issuing direction + conviction; a deterministic structurer that picks every strike, wing, DTE, and size — LLMs never compute a tradeable number. Every proposal then faces Sgt. Gate: twelve deterministic gates (sanity, regime, VRP edge, event blackout, defined-risk atomicity, liquidity, credit quality, 1% NAV position cap, 5% portfolio cap, concentration, duplicate-idempotency, and a −2% daily flatten-and-halt) — all scored, none short-circuit, every verdict journaled with reason codes. The executor (XQ) places atomic multi-leg orders through Alpaca's Trading API with idempotent client_order_ids, with the official MCP server as the agent tool surface and the Alpaca CLI providing independent position reconciliation every cycle. When a position sours, Sage's post-mortem assigns root cause and writes a boolean-checkable lesson into L3 memory that is injected into every future debate — and its parameter proposals are restrict-only, so the desk can only become more careful over time. The whole desk narrates itself in a live web UI: real-time SSE feed, live equity curve and position marks from option snapshots, the complete filterable journal, and a copilot ("invest in NVDA") that faces the same analysts, debates, and gates — including honest rejections, shown with reasons. Doctrine: LLMs argue; the math decides; Alpaca executes
Category tags: