WealthAgents

A full team of AI agents debates your finances and delivers a personalized wealth plan in 30 seconds. Your autonomous wealth management firm, in your pocket.

Comment

WealthAgents: Your Autonomous Wealth Management Firm

Inspiration

Most people will never sit across from a financial advisor. The barrier is not motivation, it's access. A single session with a certified financial planner costs $200 to $400, and the advice you get is often templated, slow, and backward-looking.

We asked a simple question: What if you could deploy an entire firm of financial experts the moment you pressed a button?

Not a chatbot. Not a calculator. A genuine multi-agent system where specialists argue, debate, and synthesize, the way a real wealth management team operates behind closed doors.

That idea became WealthAgents.

What We Built

WealthAgents is a full-stack autonomous wealth planning platform. Fill out a 2-minute intake form, and a 17-call AI agent pipeline fires in parallel and sequence, delivering a personalized financial dashboard in under 30 seconds.

The pipeline runs in 5 layers:

Layer 1 — Analyst Team (parallel) Four specialist agents analyze your profile simultaneously against live market data (FRED + Yahoo Finance):

  • Cash Flow Analyst
  • Retirement Analyst
  • Housing Analyst
  • Investment Analyst

Layer 2 — Bull vs. Bear Debate (sequential) A Bull agent and Bear agent run two adversarial debate rounds, each responding to the other's arguments. A Facilitator referee then scores the debate and produces an aggression dial $\alpha \in [0, 1]$ where:

$$\alpha = 0 \Rightarrow \text{fully conservative}, \quad \alpha = 1 \Rightarrow \text{fully aggressive}$$

Layer 3 — Portfolio Allocator Takes the debate verdict and Layer 1 reports to propose a percentage split across 5 buckets:

$$\sum_{i \in {\text{emergency, retirement, investing, house, speculative}}} w_i = 100\%$$

Layer 4 — Risk Review (parallel + sequential) Three risk agents (aggressive, neutral, conservative) stress-test the proposed allocation simultaneously, then a Risk Manager synthesizes their views.

Layer 5 — Wealth Manager Final synthesis into a full Wealth Plan with health score, headline, milestones, ETF picks, warnings, and career projections.

How We Built It

  • Backend: FastAPI with Server-Sent Events (SSE) for real-time streaming, so the frontend renders each agent's output as it arrives.
  • Agents: Fetch.ai's uAgents framework for the Layer 2 debate agents (Bull, Bear, Facilitator), registered on the Agentverse network.
  • LLM: ASI:One (Fetch.ai) as the language model powering all agent reasoning calls.
  • Market Data: Live FRED API (mortgage rates, CPI, 10yr treasury) + yfinance (trending stocks).
  • Frontend: Next.js 15 + Tailwind CSS + Recharts, with a live streaming agent trace view that transitions into a polished dashboard.
  • Auth: Email/password login with SQLite (local) or Postgres (production), plus Supabase OAuth bridge.

Challenges

Making agents argue meaningfully. Getting the Bull and Bear to genuinely respond to each other's prior arguments, not just repeat themselves, required careful prompt engineering with full transcript injection at each round.

SSE + async pipeline. Streaming 17 sequential and parallel LLM calls through a single SSE connection, without dropping events or hitting race conditions, required significant work on the asyncio event model.

Real-time UX. Designing a UI that feels alive while agents are running and transitions gracefully into a polished dashboard on completion without any page reload or state loss.

Keeping it grounded. AI agents are verbose. Every output schema was strictly typed with Pydantic, so hallucinated fields simply couldn't make it into the final plan.

What We Learned

Agentic systems are not just "LLM with more steps." The hardest part is not getting one agent to reason well, but rather designing the contracts between agents so that each layer's output is exactly what the next layer needs. Fetch.ai's uAgents framework made the messaging layer cleaner, and ASI:One gave us a model that could maintain structured financial reasoning over long, multi-turn contexts.

We also learned that the best product moment is not the output, it is the journey. Watching agents debate in real time, seeing the Bull and Bear clash, makes users trust the result far more than receiving a static PDF ever could.

Built With

Share this project:

Updates