AlphaAgents introduces a modular multi-agent framework for equity portfolio construction, leveraging Large Language Models (LLMs) to automate and enhance stock selection. The system comprises three specialized agents—Fundamental, Sentiment, and Valuation—each equipped with role-specific prompts and tools (e.g., RAG for financial reports, summarization for news, and computational tools for volatility). These agents collaborate via a group chat and engage in a debate mechanism to reach consensus, mimicking human investment committee reasoning. The framework is built on Microsoft AutoGen and uses GPT-4o as the underlying LLM.

Backtesting on 15 technology stocks from February to May 2024 shows that the multi-agent portfolio outperforms single-agent and benchmark portfolios in a risk-neutral setting, achieving higher cumulative returns and rolling Sharpe ratios. In a risk-averse setting, all portfolios underperform due to conservative stock selection during a bullish market, but the multi-agent portfolio exhibits lower drawdowns and better risk-return balance. The study also demonstrates how risk tolerance can be embedded via prompt engineering, though risk-seeking profiles were indistinguishable from risk-neutral.

The paper highlights the potential of multi-agent LLM systems to mitigate cognitive biases and improve decision quality in finance. While focused on stock selection, the framework is modular and can be extended to include additional agents (e.g., technical analysis, macro-economic) and integrated with portfolio optimization models like Mean-Variance or Black-Litterman. The transparent reasoning logs address evaluation challenges and support human-in-the-loop oversight.