Exploring Multi‑Agent Applications in Financial Scenarios and the agentUniverse Framework

At the AICon Global AI Development and Application Conference, Chen Hong, Director of Intelligent Service Algorithms at Ant Group's Wealth Insurance Business Unit, presented "Exploring Multi‑Agent Applications in Financial Scenarios," outlining recent progress and challenges of multi‑agent systems in finance and introducing Ant's self‑developed multi‑agent framework, agentUniverse.

The talk differentiates between large language models (LLMs), single agents, and multi‑agent systems, emphasizing that LLMs are stateless functions whose state must be managed externally via prompts, whereas agents become stateful machines that interact with environments through defined perception, planning, action, and memory cycles.

In financial contexts, tasks are categorized as information‑dense, knowledge‑dense, and decision‑dense. Information‑dense tasks require rapid, high‑frequency data updates and noise filtering; knowledge‑dense tasks involve conflicting expert opinions and bounded knowledge; decision‑dense tasks face uncertainty and asymmetry, demanding precise identification of critical decision points.

Standard solutions such as Retrieval‑Augmented Generation (RAG), knowledge graphs, and Chain‑of‑Thought prompting often fall short in finance because they struggle with noisy multi‑document inputs, conflicting or bounded knowledge, and uncertain, asymmetric decision making.

To address these gaps, the proposed approach assigns rigorous knowledge injection to the LLM layer while delegating process‑oriented, domain‑specific Know‑How to specialized agents, forming a two‑layer system that balances rigor and expertise.

The agentUniverse open‑source framework implements this vision. Its core PEER (Plan‑Execute‑Express‑Review) paradigm orchestrates four distinct agents: a planner that decomposes queries, an executor that performs retrieval or computation, an expresser that synthesizes results, and a reviewer that validates output before final release.

Using agentUniverse, Ant has built a financial research assistant (投研支小助) that supports report interpretation, market analysis, policy impact assessment, and macro‑economic studies. In internal tests, the assistant helps a research analyst produce over 100 high‑quality reports and analyses per day, covering more than 50 financial events.

Examples include automated analysis of Nvidia’s FY2024 Q4 earnings for AI industry outlook, structured policy impact evaluations, and macro‑economic briefings, all generated through nested PEER workflows that incorporate expert‑derived frameworks.

Future work will publish a paper detailing the PEER framework and its effectiveness in complex financial tasks.