Multi-Agent Financial Research Assistant
An LLM-orchestrated multi-agent system that automates financial research pipelines: Data Collection Agent → Quantitative Analysis Agent → Report Generation Agent, with tool use and context memory — compressing research cycles from days to minutes.
Overview
Key Features
Supervisor-Worker architecture: Supervisor handles task decomposition, sub-agent scheduling, and result aggregation with dynamic routing
Tool calling (Function Calling): each agent has a dedicated toolset including financial data APIs, calculators, and chart generators
Context memory management: short-term state storage with history summarization to prevent token overflow
Parallel agent execution: data collection and initial analysis run concurrently, reducing total latency by ~60%
Human-in-the-loop: critical nodes (risk rating, conclusion sign-off) pause for human approval, ensuring output trustworthiness
Methodology
Built on LangGraph StateGraph. Each node is an agent; edges are conditional routing rules. The Supervisor uses ReAct (reason → act → observe) planning loops. Worker agents return structured JSON via Pydantic models. The Supervisor aggregates outputs and passes them to the Report Agent, which produces a Markdown report with sections: Executive Summary, Financial Analysis, Risk Assessment, and Investment Thesis.
Tech Stack
Project Info
Multi-Agent Pipeline Demo
Select a research topic, launch the 4-layer agent pipeline, and watch tool call logs and final report generation in real time.
Financial Research Multi-Agent System
4-layer agent pipeline with tool call logs and structured output.
0/4