
The integration of Large Language Models (LLMs) into algorithmic trading presents significant potential, yet pure LLM systems frequently hallucinate numerical computations and lack rigorous risk management. We present OrbiTrade, an autonomous, multi-agent LLM system designed specifically for executing options strategies. The core philosophy of this architecture is the strict separation of deterministic quantitative computation — utilizing the Black-Scholes model, the Greeks, and the Kelly Criterion — from qualitative reasoning, thereby preventing pricing and risk hallucinations. Orchestrated via LangGraph, the system operates through a specialized four-agent pipeline: each agent focuses on a different side of the decision, and each hands the next a structured, validated artifact rather than free text. Furthermore, OrbiTrade achieves continuous improvement without model retraining through a post-trade reflection loop that logs trade outcomes and extracts actionable lessons for future contextual recall.
4 Sep 2026