How the Multi-Agent System in AI Berkshire Works: Three-Layer Architecture Explained

AI Berkshire implements a multi-agent system using a three-layer architecture that orchestrates four specialized research agents to run in parallel, enforced by deterministic financial tools and aggregated by a human-led team lead.

The xbtlin/ai-berkshire repository delivers a sophisticated multi-agent research framework designed for institutional-grade investment analysis. At its core, the Investment-Team skill instantiates a miniature research team that mirrors the decision-making processes of legendary investors like Warren Buffett and Charlie Munger. This architecture separates concerns into distinct layers while ensuring rigorous data validation through deterministic Python utilities.

Three-Layer Architecture Overview

AI Berkshire organizes its multi-agent system into three discrete layers that separate user interaction from execution and calculation.

• Skill Layer: Defines high-level commands such as /investment-team that map user requests to concrete workflows. The command definition and execution steps live in skills/investment-team.md.

• Agent Layer: Instantiates four independent agents, each embodying a master-investor perspective. These agents execute simultaneously, performing their own web searches, data validation, and analysis before reporting back to a central orchestrator.

• Tool Layer: Supplies deterministic utilities via tools/financial_rigor.py that agents must invoke instead of "hand-calculating." This layer provides financial-rigor checks, cross-validation, and report-audit capabilities.

The Four Specialized Agents

The Investment-Team skill constructs a research team of four agents, each assigned a specific analytical dimension and role model:

• Business Analyst (段永平 perspective): Evaluates business models, competitive moats, and user value propositions.
• Financial Analyst (巴菲特 perspective): Performs three-scenario valuation, ROE analysis, and market-cap verification.
• Industry Researcher (芒格 perspective): Analyzes industry trends, competitive positioning, and long-term market dynamics.
• Risk Assessor (李录 perspective): Identifies downside scenarios, regulatory risks, and counter-arguments to the investment thesis.

Each agent receives a role-specific prompt that mandates the use of WebSearch and financial-data standards, ensuring consistent research methodology across the team.

Execution Workflow

The multi-agent system follows a rigorous seven-step execution pattern defined in skills/investment-team.md (lines 34-106).

Team Creation and Task Definition

The workflow begins with a TeamCreate call that establishes a named research unit:

{
  "action": "TeamCreate",
  "team_name": "meituan-research",
  "agent_type": "team-lead"
}

Following team creation, the system generates four TaskCreate calls, one for each analyst role. Each task contains a specific subject and detailed description that enforces data-source rules and required analysis dimensions (business model, financials, industry, risk) according to skills/investment-team.md (lines 42-92).

Parallel Execution and Background Processing

All four Task objects dispatch simultaneously in a single message using the Task tool with run_in_background: true:

{
  "action": "Task",
  "tasks": [
    {"name": "business-analyst", "prompt": "..."},
    {"name": "financial-analyst", "prompt": "..."},
    {"name": "industry-researcher", "prompt": "..."},
    {"name": "risk-assessor", "prompt": "..."}
  ],
  "run_in_background": true
}

This parallelism multiplies search coverage and analytical viewpoints, delivering research depth comparable to a real investment team while maintaining workflow reproducibility.

Data Validation and Tool Integration

Inside each agent's prompt, the system highlights the rule: "Financial data must come from two independent sources." Agents must call Python utilities from tools/financial_rigor.py to embed verification results directly in their partial reports. For example:

python3 ~/ai-berkshire/tools/financial_rigor.py verify-market-cap --ticker 3690.HK

The financial_rigor.py module uses decimal.Decimal for all numeric calculations, avoiding floating-point drift and ensuring deterministic outputs across different execution environments.

Progress Reporting and Aggregation

Agents transmit interim findings to the team lead via SendMessage. The lead displays a live progress table summarizing each agent's completion status and top 3-5 insights. After all four reports arrive, the lead issues a shutdown_request to each agent and aggregates the content into a final structured research report containing:

1. A single-sentence investment conclusion
2. A four-dimensional rating table
3. A data snapshot with verification timestamps
4. Bull-vs-bear arguments
5. A compliance checklist

The final report passes through report_audit.py for random-sample verification before being written to a dated markdown file. The system then removes team resources with TeamDelete.

Architectural Safeguards and Quality Controls

AI Berkshire incorporates three specific mechanisms to ensure research quality:

Information-Richness Rating: Before execution, the system assigns an A/B/C rating that determines how aggressively agents probe for counter-evidence. This rating influences research strategy and is presented to agents in their initial prompts (lines 21-29 of skills/investment-team.md).

Bias Mitigation: The team lead must verify whether any agent's output simply mirrors market consensus. If detected, the lead forces a "reverse-check" to surface hidden risks and contrarian viewpoints.

Deterministic Tooling: All numeric calculations route through tools/financial_rigor.py, which enforces financial-rigor standards including market-cap verification and three-scenario valuation. This prevents calculation errors that could propagate through the multi-agent workflow.

Summary

• AI Berkshire uses a three-layer architecture (Skills, Agents, Tools) to separate user commands from parallel execution and deterministic calculation.
• The Investment-Team skill orchestrates four specialized agents (Business, Financial, Industry, Risk) that run simultaneously using run_in_background: true.
• Agents must invoke deterministic utilities from tools/financial_rigor.py for all financial calculations, using decimal.Decimal to ensure precision.
• The workflow enforces dual-source verification for all financial data, with automated auditing via report_audit.py before final publication.
• Bias-mitigation protocols require the team lead to challenge consensus views and force reverse-checks when agents echo market sentiment.

Frequently Asked Questions

What is the three-layer architecture in AI Berkshire's multi-agent system?

The three-layer architecture separates concerns into Skills (user-facing commands like /investment-team), Agents (the four parallel research specialists), and Tools (deterministic Python utilities in tools/financial_rigor.py). This separation ensures that agents focus on analysis while rigorous calculation and data validation happen through standardized, reproducible utilities.

How do the four agents in the Investment-Team skill collaborate?

The four agents (Business Analyst, Financial Analyst, Industry Researcher, and Risk Assessor) work in parallel rather than sequentially, each receiving role-specific prompts that mandate independent research. They report findings to a human-led team lead via SendMessage, and the lead aggregates these into a unified report only after all four agents complete their analysis and pass data-verification checks.

What ensures data accuracy in the AI Berkshire multi-agent system?

Data accuracy relies on mandatory dual-source verification enforced through skills/financial-data.md standards. Agents must call tools/financial_rigor.py functions (such as verify-market-cap) rather than performing calculations independently. The system uses decimal.Decimal for all numeric operations to prevent floating-point errors, and final reports undergo random-sample verification via report_audit.py before publication.

How does the system handle bias and consensus thinking?

The team lead actively monitors for consensus bias by checking whether agent outputs merely reflect market sentiment rather than independent analysis. When detected, the lead triggers a "reverse-check" that forces agents to seek out contrarian evidence and hidden risks. Additionally, the information-richness rating (A/B/C) adjusts research depth based on the need for counter-evidence investigation.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}