Someone just created a hedge fund model using AI agents. A rapidly trending open-source Python project called “Trading Agents” simulates an entire hedge fund using multiple LLM-powered agents. Instead of a single model making decisions, it mirrors how real trading firms operate—with specialized roles that analyze, debate, and approve trades collaboratively. The framework breaks down a trading firm into distinct AI agents: 1. Analyst Layer (parallel execution) - Fundamentals analyst → financials & valuation - Sentiment analyst → social/media mood - News analyst → macro + breaking events - Technical analyst → indicators (MACD, RSI, etc.) Each produces independent reports—disagreement is intentional and valuable. 2. Research Layer (debate engine) - Bull agent vs. Bear agent - They argue over multiple rounds using analyst data - Designed to surface stronger, more defensible conclusions 3. Decision Layer - Trader agent → decides timing & position size - Risk management → evaluates volatility/liquidity - Portfolio manager → final approval/rejection If approved → simulated trade executesIf rejected → system logs reasoning Unlike traditional trading systems: - Not purely rule-based - Not a black-box ML model Instead, it’s: - Transparent & auditable - Every decision is traceable: Analyst reports Debate transcripts Trade rationale Approval/rejection logic Technical Backbone - Built in Python using LangGraph (agent orchestration) - Each agent = node in a decision graph - Features: Checkpointing (resume runs) Persistent decision logs Self-reflection loop: Tracks past trades Compares performance vs. benchmarks Feeds insights back into future decisions Recent Updates (v0.2.4) - Structured outputs using schemas (more reliable decisions) - Expanded model support (OpenAI, Gemini, Claude, DeepSeek, etc.) - Docker support - Standardized rating system: Buy / Overweight / Hold / Underweight / Sell

A decent demo video of how to use Opus 4.5 to build n8n workflows.