AgentUniverse: An Enterprise‑Grade Multi‑Agent Framework for Complex Financial Analysis

Traditional large language models (LLMs) often struggle with complex, multi‑step reasoning tasks such as financial analysis, producing vague or factually inaccurate results. To address these shortcomings, Ant Group proposes the AgentUniverse framework, which enables multiple specialized agents to cooperate and collectively solve intricate problems.

AgentUniverse supplies a pattern factory that lets developers create custom multi‑agent collaboration modes while also offering ready‑made components for building individual agents. The framework’s flagship collaboration pattern, PEER, combines hierarchical collaboration and teamwork by assigning four distinct roles: Planning, Executing, Expressing, and Reviewing.

In the Planning stage, a leader agent decomposes a high‑level query (e.g., “Why did Warren Buffett reduce his BYD holdings?”) into 5‑10 sub‑questions. The Executing agents retrieve relevant information using tools such as web search and domain‑specific knowledge bases. The Expressing agent formats the gathered insights into a coherent, structured answer, while the Reviewing agent evaluates the output against predefined scoring criteria and decides whether further iteration is needed.

Applying this workflow to the Buffett case study, the PEER mode generated a detailed, multi‑paragraph analysis that outperformed a single‑agent baseline. Experimental results on the GAME24 benchmark showed that PEER raised accuracy from 84 % to 99 % (b=1) and reduced the average number of retry rounds, demonstrating superior efficiency and reliability for multi‑step reasoning tasks.

Beyond this example, PEER consistently achieved strong performance on other reasoning suites such as Folio Wiki, MATH, and Titanic. Within Ant Group, the framework powers the “投研支小助” AI assistant, boosting analysts’ productivity by over 50 % and achieving a 70 % adoption rate.

AgentUniverse is open‑source; Python developers can install it via pip install agentUniverse and access the repository at https://gitee.com/agentUniverse/agentUniverse. The project invites developers and industry partners to explore the limitless possibilities of AI multi‑agent collaboration.