Multi-Agent Parallel Research Team in AI Berkshire: Architecture and Implementation

AI Berkshire's multi-Agent parallel research team coordinates four specialized AI agents simultaneously to conduct deep, balanced investment analysis, eliminating single-agent bias while accelerating research turnaround through concurrent task execution.

The AI Berkshire repository implements a sophisticated multi-agent architecture designed to automate institutional-grade equity research. At its core, the multi-Agent parallel research team dynamically assembles for each target company ({公司名}) to provide structured, opinion-diverse coverage across all critical investment pillars.

What is the Multi-Agent Parallel Research Team?

The multi-Agent parallel research team is an on-the-fly construct created specifically for each investment target. According to the canonical definition in skills/investment-team.md, the system spawns a team-lead agent plus four specialized research agents that execute concurrently. This architecture ensures that no single perspective dominates the analysis, while the parallel execution model dramatically reduces research latency compared to serial workflows.

The team operates through distinct analytical lenses derived from value investing legends: the business-analyst adopts the 段永平 (Duan Yongping) framework, the financial-analyst applies 巴菲特 (Buffett) rigor, the industry-researcher utilizes 芒格 (Munger) mental models, and the risk-assessor employs 李录 (Li Lu) risk frameworks. The team-lead synthesizes these inputs through the 四大师 (Four Masters) combined view.

The Four Specialized Research Agents

Each agent in the parallel research team maintains a distinct analytical focus, ensuring comprehensive coverage of the investment thesis.

Business Analyst (段永平 Lens)

The business-analyst agent evaluates the target company's business model, competitive moats, and user value proposition. This agent focuses on qualitative factors that determine long-term economic sustainability, applying the investment philosophy of 段永平 to assess whether the company possesses durable competitive advantages.

Financial Analyst (巴菲特 Lens)

The financial-analyst agent conducts deep quantitative analysis of financial statements, profitability metrics, and valuation models. Following the 巴菲特 (Buffett) analytical lens, this agent enforces financial-rigor verification using validation tools from tools/financial_rigor.py to cross-check market-cap calculations and valuation assumptions.

Industry Researcher (芒格 Lens)

The industry-researcher agent maps the competitive landscape and industry structure. Through the 芒格 (Munger) lens of multidisciplinary thinking, this agent identifies industry trends, competitive dynamics, and structural factors that influence the company's long-term positioning.

Risk Assessor (李录 Lens)

The risk-assessor agent evaluates investment risks and management quality. Applying the 李录 (Li Lu) framework of margin-of-safety and management assessment, this agent identifies red flags, governance issues, and potential downside scenarios that could impair the investment thesis.

The 8-Step Parallel Research Workflow

The workflow defined in skills/investment-team.md orchestrates the team through a rigorous, reproducible pipeline:

1. Framework Assessment – Display the analytical framework and assess information richness (A/B/C rating).

2. Team Creation – Instantiate TeamCreate with the naming convention {公司名}-research.

3. Task Definition – Create four detailed tasks covering commercial, financial, industry, and risk dimensions.

4. Parallel Launch – Dispatch all four TaskCreate calls simultaneously in a single message, enabling concurrent execution.

5. Progress Collection – The team-lead monitors status tables and extracts key takeaways as sub-reports arrive.

6. Agent Shutdown – Terminate the four research agents cleanly once all deliverables are complete.

7. Master Synthesis – Generate a comprehensive report combining the four analytical dimensions, including a one-sentence thesis, scoring matrix, risk checklist, and final recommendation.

8. Audit and Archive – Execute report_audit.py to validate outputs and archive the research package.

Implementation: Creating and Running the Research Team

Step 1: Initialize the Research Team

Create the team structure using the OpenCopilot CLI:

# Replace <company> with the target ticker or name, e.g. "meituan"

team_name="${company}-research"
opencopilot task create \
  --team-name "$team_name" \
  --agent-type team-lead \
  --subject "启动 AI Berkshire 投研团队"

Step 2: Define the Four Parallel Tasks

Create concurrent tasks for each specialized agent:

# Business analyst task

opencopilot task create \
  --team-name "$team_name" \
  --subagent-type general-purpose \
  --name business-analyst \
  --subject "分析${company}商业模式、护城河与用户价值" \
  --description @tasks/business_analyst.md

# Financial analyst task

opencopilot task create \
  --team-name "$team_name" \
  --subagent-type general-purpose \
  --name financial-analyst \
  --subject "分析${company}财务数据、盈利能力与估值" \
  --description @tasks/financial_analyst.md

# Industry researcher task

opencopilot task create \
  --team-name "$team_name" \
  --subagent-type general-purpose \
  --name industry-researcher \
  --subject "分析${company}所在行业格局与竞争态势" \
  --description @tasks/industry_researcher.md

# Risk assessor task

opencopilot task create \
  --team-name "$team_name" \
  --subagent-type general-purpose \
  --name risk-assessor \
  --subject "评估${company}投资风险与管理层质量" \
  --description @tasks/risk_assessor.md

Step 3: Launch Tasks in Parallel

Use Python to batch-create tasks simultaneously via the API:

import json, subprocess, os

team = f"{company.lower()}-research"
tasks = [
    {"name":"business-analyst","subject":"商业模式分析"},
    {"name":"financial-analyst","subject":"财务与估值分析"},
    {"name":"industry-researcher","subject":"行业与竞争分析"},
    {"name":"risk-assessor","subject":"风险与管理层评估"},
]

payload = {
    "team_name": team,
    "tasks": [
        {
            "subagent_type":"general-purpose",
            "run_in_background":True,
            "name":t["name"],
            "subject":t["subject"],
            "description":f"请完成任务 {i+1}: {t['subject']}"
        }
        for i,t in enumerate(tasks)
    ]
}

# Send a single request to the OpenAI-function endpoint

subprocess.run(["opencopilot", "task", "bulk-create"], input=json.dumps(payload).encode())

Step 4: Synthesize and Audit Results

Collect completed reports and generate the final analysis:

# List completed sub-reports

opencopilot team status "$team_name"

# Download raw reports when all four are marked 'completed'

opencopilot team download "$team_name" --output ./reports/${company}_research_raw.zip

# Execute final synthesis and validation

python3 tools/report_audit.py extract --report ./reports/${company}_final.md

Key Files and Tools

The multi-agent system relies on several critical components:

• skills/investment-team.md – Canonical definition of the four-agent parallel research framework and workflow steps.
• tools/financial_rigor.py – CLI utilities used by the financial-analyst agent for market-cap verification and valuation cross-checks.
• tools/report_audit.py – Post-research validation and audit tool that ensures output quality before archiving.
• codex-prompts/investment-team.md – Synchronized prompt version for Codex compatibility.
• scripts/sync-codex-skills.py – Maintenance script that keeps the Codex prompt in sync with the canonical skill definition.

Summary

The multi-Agent parallel research team in AI Berkshire delivers several critical advantages:

• Eliminates single-agent bias by requiring four independent analytical perspectives before synthesis.
• Ensures comprehensive coverage across business model, financials, industry structure, and risk dimensions.
• Accelerates research velocity through true parallel execution rather than serial task processing.
• Enforces methodological rigor via mandatory cross-validation steps embedded in each sub-task and final audit requirements.
• Maintains reproducibility through canonical skill definitions that standardize analysis for every company investigated.

Frequently Asked Questions

How does the multi-agent approach reduce bias compared to single-agent analysis?

The parallel architecture mandates that four independent agents—each operating through distinct investment philosophies (段永平, 巴菲特, 芒格, 李录)—complete their analysis before any synthesis occurs. This structural requirement prevents confirmation bias and ensures that contrarian viewpoints surface during the research process, as no single agent can dominate the final output without peer validation.

What triggers the creation of a research team for a specific company?

The system dynamically instantiates a new team using the TeamCreate operation whenever a user initiates research on a target company ({公司名}). The team-lead agent evaluates the information richness (A/B/C rating) of available data before formally launching the four parallel research streams, ensuring that sufficient context exists to support meaningful analysis.

How does the team ensure data rigor across parallel streams?

Each specialized agent incorporates domain-specific validation protocols. The financial-analyst specifically invokes utilities from tools/financial_rigor.py to cross-verify market-cap calculations and valuation multiples. Additionally, the final report_audit.py step validates the synthesized master report against the four source documents before archiving, ensuring consistency and accuracy across all dimensions.

Can the workflow be customized for different industries or investment strategies?

While the current implementation in skills/investment-team.md standardizes on the four value-investing lenses (段永平, 巴菲特, 芒格, 李录), the modular task structure allows for customization. Users can modify the task descriptions in @tasks/business_analyst.md and related files, or extend the scripts/sync-codex-skills.py pipeline to generate variant teams for growth investing, credit analysis, or sector-specific deep dives while maintaining the parallel execution architecture.

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