“"The pattern of specialized agents debating to reach better decisions is the feature everyone will be building in 2025." — Yuval Avidani, yuv.ai/blog/ai-hedge-fund-building-investment-teams-from-ai-agents (published Jan 5, 2026, fetched 2026-05-10)”
You know that feeling when you want to study how multiple AI agents collaborate on a decision, but every example in the docs is trivially simple — one agent, one task, no shared state to manage? Building a realistic multi-agent system means designing state passing, handling conflicting agent outputs, wiring LLM calls into a graph, streaming results to a UI, and adding risk controls — all before you can observe anything interesting. This repo gives you a working 19-agent system with investor personas, risk limits, and a React flow editor so you can study the orchestration patterns without building the plumbing. The financial domain also makes the agent outputs immediately legible: buy, hold, or sell is a clearer output than abstract task completion.
You open a browser, drag investor-persona agents and analysis agents onto a canvas, and connect them into a flow — deciding which analysts feed into the risk manager before the portfolio manager makes the final call. The frontend sends that graph definition via HTTP to a FastAPI backend, which resolves node identifiers to registered Python functions and hands the topology to LangGraph. LangGraph runs each agent sequentially: every agent reads the shared immutable AgentState (portfolio positions, cash balances, previous signals, market data), runs its analysis through an LLM call, and writes its verdict back to the state. The risk manager applies hard position limits after all analysts complete, then the Portfolio Manager issues the final recommendation. Results stream to your browser in real time via Server-Sent Events — you watch each agent's reasoning appear as it finishes.
If you're building with LangGraph or LangChain and want to study a real multi-agent orchestration pattern beyond toy examples, this repo shows you 19 coordinated agents with shared state management, SSE streaming, and a React visual editor already wired together. Also useful if you're prototyping fintech AI tooling and need a reference architecture for agent-based analysis pipelines. Not appropriate if you need production trading infrastructure — the repo has no declared license (legal ambiguity for commercial derivatives), no backtesting benchmarks showing agent outputs outperform a baseline...
Worth exploring if you're learning LangGraph multi-agent patterns or building a demo of agent-based financial analysis — the architecture is well-structured, and the React + FastAPI + LangGraph stack is a clean reference for this class of system. Not ready for production use: no declared license creates legal ambiguity for commercial derivatives, no performance benchmarks exist showing the agents outperform a market baseline, ETF data integration is broken (issue #617), and LLM calls can hang indefinitely without timeouts (issue #618).
Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.