Vanguard_Sentinel_options_trading_agent

Vanguard Sentinel is an autonomous options-trading agent built for the Alpaca AI Trading Agents Hackathon, designed around one core principle: risk limits are enforced in deterministic code, never left to an AI's discretion. The agent continuously scans a basket of twelve liquid ETFs, evaluating implied volatility against realized volatility, price efficiency, and macro-event risk to classify each instrument's market regime in real time. Depending on conditions, it selects between selling premium via iron condors, buying volatility via long strangles ahead of scheduled catalysts, or standing aside entirely — and standing aside was, by far, its most common decision. Every trade candidate that survives Vanguard Sentinel's deterministic risk gates — a hard equity floor, per-trade risk caps, defined-risk-only structures, and position concentration limits — still has to win an internal debate. A Bull agent argues for entry, a Bear agent argues against it, and a Judge model renders the final verdict, with every transcript logged and publicly auditable. Over a multi-day live paper-trading run, the agent evaluated [N] trade proposals and executed [N], with the large majority rejected automatically before any AI reasoning even occurred. It ended the session at [final equity], a [P&L%] result, having never approached its hard safety floor. Mid-run, a real multi-leg order execution bug was discovered and fixed live — a redesign to atomic order submission that's part of the public commit history, not hidden from it. Built on Alpaca's paper trading API and MCP server, with a dual-provider LLM reasoning layer (Featherless AI primary, local Ollama fallback) ensuring the agent never trades blind even if one AI provider fails