Frameworks

What Is an AI Investment Committee? Four Agents Argue About Value Investing for You

An open-source project encodes Buffett, Munger, Duan Yongping, and Li Lu's investment frameworks into a multi-agent system that runs inside Claude Code. Here's how it works—and where it falls short.

Font size

ai-berkshire is an open-source project by xbtlin on GitHub, pitched as "Berkshire Hathaway for the AI era." The premise is straightforward: take the decision frameworks of four value investing masters—Buffett, Munger, Duan Yongping, and Li Lu—and encode them into a multi-agent system that runs inside Claude Code.

Sounds gimmicky. After digging through the repo, it turns out to be more serious than most AI investing toys.

The Problem It Solves

Retail investors reading prospectuses or 10-Ks either can't finish them or can't extract anything useful. Most AI investing tools on the market are single-LLM summaries—you feed in a financial report and get back a "comprehensive analysis." Done.

What's missing is multi-perspective adversarial analysis.

ai-berkshire explicitly encodes each master's decision framework into prompts plus checklists, then has the agents challenge each other, simulating a real investment committee. This is a different animal from asking ChatGPT whether to buy a stock.

How to Use It

Minimal working example inside the Claude Code CLI:

# 1. Clone the repo
git clone https://github.com/xbtlin/ai-berkshire.git
cd ai-berkshire

# 2. Load four agent configs in Claude Code
# .claude/agents/buffett.md  - Buffett lens (ROIC + moat)
# .claude/agents/munger.md   - Munger lens (mental models)
# .claude/agents/duan.md     - Duan Yongping lens (right business + right people)
# .claude/agents/li.md       - Li Lu lens (deep value + long-term)

# 3. Trigger multi-agent adversarial research
/research --ticker=600519.SH --adversarial=true

# 4. Sample output (simplified):
# [Buffett Agent] 5-yr avg ROIC 32%, strong brand moat, meets criteria
# [Munger Agent]  challenge: Does the moat hold when the baijiu cycle turns down?
# [Duan Agent]    Management is stable, but channel reform adds uncertainty
# [Li Agent]      Current valuation is reasonable to low, long-term hold acceptable
# [Synthesis]     Overall rating: Buy (75% confidence)

No Python environment needed—just the Claude Code CLI. The repo is mostly Markdown prompts and YAML config; Python is minimal.

How It Works (Architecture)

Three layers:

(1) Master Method Layer — One agent per master, with decision checklists hardcoded into the prompt. Buffett focuses on ROIC and moat. Munger runs mental models. Duan checks for right business and right people. Li Lu looks at deep value and long-term holding. These aren't vague "act as Buffett" prompts; they're structured decision nodes.

(2) Adversarial Layer — Agents challenge each other, forced to output "I disagree with X on point Y." This is the core of the framework: not four agents stating their views independently, but actively pushing back on each other.

(3) Synthesis Layer — Aggregates multiple opinions into a final rating with a confidence score.

Flow: ticker goes in → four agents run in parallel → cross-challenge happens → summary report comes out.

What It Does Well / Poorly

Strengths:

  • Not another web UI wrapper. It plugs directly into Claude Code's workflow—running analysis feels like running code.
  • Prompts are structured checklists, not generic "you are Buffett, analyze this stock."
  • The research flow is templated and modifiable—you can add a "Cathie Wood" agent or swap in your own framework.

Weaknesses:

  • All four agents are essentially Claude, so viewpoints may homogenize. Real adversarial diversity is limited.
  • No backtesting module—historical accuracy can't be verified.
  • Support for A-shares and HK stocks depends entirely on the LLM's own ability; no specialized data source is connected.
  • 15,796 stars looks impressive, but value investing is a minority pursuit. Real activity is probably lower.

I Tried It

I ran the demo flow in Claude Code, loading agents/buffett.md and agents/munger.md to analyze a consumer stock (let's say Kweichow Moutai).

The Buffett agent focused on sustained ROIC above 25% and the brand moat. The Munger agent immediately challenged: "Does the brand moat hold when the baijiu cycle turns down?"

The whole flow took five minutes and produced a roughly 3,000-word report. The quality was noticeably higher than a single-LLM summary—at minimum, you can see two distinct thinking frameworks colliding.

But the adversarial depth is limited. If two agents share the same underlying model, disagreements tend to be tonal rather than substantive. That's my biggest reservation.

Who Should Use This

Good fit:

  • Developers already using Claude Code or Codex CLI
  • Value investors who want to systematize the logic behind their holdings
  • Beginners who need a quick way to build a research framework

Bad fit:

  • Short-term traders (this framework is long-term oriented)
  • Anyone without a Claude Code subscription
  • People who reject value investing as a methodology (no tool will help there)

Editorial Take

What's interesting here isn't "can AI trade stocks." It's the act of explicitly encoding expert decision frameworks.

The past decade has produced plenty of "let AI do X" projects, most of which amount to throwing a task at an LLM and hoping for the best. ai-berkshire's approach is closer to a cognitive architecture—taking Buffett checking ROIC or Munger running mental model checklists, that tacit knowledge, and making it explicit and executable. The real battleground for this idea isn't investing. It's domains where expert judgment is expensive: medical diagnosis, legal due diligence, code review.

What I'm unsure about: running four "master" agents on the same Claude instance—how much real adversarial tension can that actually produce? My guess is that adversarial depth depends on underlying model diversity. If future versions can route queries across GPT, Claude, and Gemini, the framework would truly take off. Otherwise it's closer to one person playing four roles—form over substance.

The other missing piece is backtesting. An investment committee that can't tell you how its recommended picks performed over the last three years has an empty confidence number (that 75%). That's the biggest gap the project needs to fill.

Source: github.com/xbtlin/ai-berkshire

Source: https://github.com/xbtlin/ai-berkshire