I built a 9-agent AI investment committee, the debate every stock sequentially - each analyst reads all previous report before writing their own

For the past few weeks I've been building an AI-powered investment research tool. Here's how it works and what I learned.

**The problem I wanted to solve**

Asking a single AI "should I buy XYZ?" gives you a vague, overly optimistic answer. There's no adversarial pressure, no one challenging the bull case.

**What I built?**

A sequential committee of 9 specialized AI analysts. Each one reads every previous report before writing their own - so later agents can challenge earlier ones.

The pipeline:

1. **Data Scout** \- live web search for current price, EPS vs consensus, analyst targets, breaking news
2. **Macro Strategist** \- Fed policy, business cycle, sector vs index
3. **Data Hunter** \- P/E, EV/EBITDA, FCF yield, ROIC, insider ownership
4. **Sentiment Analyst** \- short interest, 13F changes, insider transactions
5. **The Bear** \- hardwired to find reasons NOT to buy
6. **The Chartist** \- MA20/50/200, RSI, MFI, Fibonacci levels, entry point
7. **Devil's Advocate** \- attacks blind spots in every previous report
8. **The CIO** \- reads all 7 analysts, delivers verdict + 1–10 scorecard across 5 dimensions
9. **Portfolio Manager** \- position sizing, DCA tranches with specific prices, stop loss, two targets

**What surprised me**

The Bear and Devi's Advocate improve output quality. Without adversarial agents, the committee was too bullish. Forcing two agents to attack the thesis surface risks I wouldn't have thought to ask about.

**Technical aspects**

* single HTML file, runs in the browser
* Anthropic API (Haiku for 7 agents, Sonnet fora CIO and Devil's Advocate)
* Live web search via Anthropic's web search tool
* Privacy - no sever, no data leaves your device
* \~$ 0.10 per full analysis

**Is it worth building a tool like this? Do you guys use anything similar?**