# How the AI Berkshire /investment-team Skill Manages Multi-Agent Parallel Research with Distinct Perspectives

> Discover how the AI Berkshire /investment-team skill manages multi-agent parallel research using four distinct agents. Learn about its data verification and output aggregation for unified investor dossiers.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: how-to-guide
- Published: 2026-06-26

---

**The AI Berkshire `/investment-team` skill orchestrates multi-agent parallel research by deploying four specialized sub-agents simultaneously—each embodying a distinct investment philosophy—while enforcing rigorous data-verification standards and aggregating outputs into a unified investor dossier.**

The `xbtlin/ai-berkshire` repository implements this architecture in [`skills/investment-team.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/investment-team.md), leveraging Claude Code’s **Team** and **Task** primitives to simulate a four-person research team with complementary analytical frameworks. This design enables true parallelism where each agent investigates the target company through a specific lens without sequential bottlenecks.

## Team Composition: Four Distinct Investment Perspectives

According to the source code in [`skills/investment-team.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/investment-team.md) (lines 11-18), the skill initializes a structured research team comprising five distinct roles. The **team-lead** (user) coordinates the effort, while four AI agents adopt specific investment philosophies:

- **Business Analyst**: Embodies the 段永平 (Duan Yongping) view, focusing on business model quality and competitive moats.
- **Financial Analyst**: Adopts the 巴菲特 (Warren Buffett) perspective, concentrating on financial metrics and intrinsic valuation.
- **Industry Researcher**: Channels the 芒格 (Charlie Munger) approach, emphasizing industry dynamics and competitive landscape.
- **Risk Assessor**: Applies the 李录 (Li Lu) framework, evaluating management integrity and downside risks.

This multi-agent parallel research structure ensures that no single analytical bias dominates the final output.

## Pre-Execution: AI-Bias Rating and Team Initialization

Before spawning agents, the skill evaluates the **AI researchability** of the target company, assigning an A-B-C rating that measures information richness and potential AI bias (lines 21-30). This rating propagates to all sub-agents via the task descriptions, allowing each to adjust its confidence level and research depth accordingly.

The orchestration begins with a **TeamCreate** call (lines 34-40), which instantiates the team with a company-specific identifier and assigns the user as the *team-lead*.

## Parallel Task Distribution and Execution

The core of the multi-agent parallel research capability lies in four simultaneous **TaskCreate** operations defined in lines 40-92. Each task targets a dedicated research domain:

1. **Business Model Analysis**: Moat durability and competitive advantages.
2. **Finance & Valuation**: Financial statements and intrinsic value calculations.
3. **Industry & Competition**: Market positioning and secular trends.
4. **Risk & Management**: Governance quality and downside scenario modeling.

Rather than executing sequentially, the skill launches all four agents in a single message using the **Task** tool with parameters `subagent_type: general-purpose` and `run_in_background: true` (lines 94-104). Each agent receives a prompt template injecting the company name, specific role identifier, and tailored task description containing perspective-specific analytical frameworks.

## Data Verification and Cross-Validation

To maintain rigorous standards across parallel streams, each agent must adhere to a strict data-sourcing contract. The skill mandates **two independent sources** for all financial data—for example, cross-referencing macrotrends against stockanalysis for US equities.

Additionally, agents invoke the [`financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/financial_rigor.py) utility (lines 64-68) to verify market-cap calculations, valuation multiples, and scenario modeling. This ensures that despite operating independently, all four perspectives rely on consistently validated inputs defined in [`skills/financial-data.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/financial-data.md).

## Aggregation and Final Synthesis

While sub-agents execute in parallel, the team-lead monitors progress through a live tracking table (lines 30-34). Upon completion of all four reports, the lead sends `shutdown_request` signals to each sub-agent and compiles the final deliverable.

The synthesized report includes:
- Executive summary integrating all four perspectives.
- Scoring table with weighted metrics.
- Data snapshot with verified sources.
- Pros/cons matrix.
- Investment checklist.
- Final recommendation.

The completed dossier is written to a dated markdown file, processed through [`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py) for random-sample verification, and the entire team structure is dismantled via **TeamDelete** (lines 84-102).

## Invoking the Investment Team

To trigger the multi-agent parallel research workflow from Claude, use the following command syntax:

```markdown

# Example: invoking the investment‑team skill from Claude

/investment-team 公司名="腾讯" 日期=20260630

```

For programmatic integration, construct a payload targeting the Claude API:

```python
import json, subprocess

# Build the skill invocation payload

payload = {
    "skill": "/investment-team",
    "arguments": {
        "公司名": "腾讯",
        "日期": "20260630"
    }
}

# Send to Claude (the surrounding platform handles the HTTP call)

response = subprocess.check_output(
    ["curl", "-X", "POST", "-d", json.dumps(payload),
     "https://api.anthropic.com/v1/complete"], text=True)

print("Claude response:", response)

```

## Summary

- The **/investment-team** skill creates four specialized agents (Business, Financial, Industry, Risk) plus a human team-lead, each adopting distinct investment philosophies (段永平, 巴菲特, 芒格, 李录).
- **Parallel execution** is achieved through `run_in_background: true` and `subagent_type: general-purpose` flags in the **Task** tool, launching all research streams simultaneously.
- An **AI-bias rating** (A-B-C) calibrates agent confidence before research begins, ensuring appropriate skepticism for information-sparse targets.
- **Data rigor** is enforced via mandatory dual-source verification and the [`financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/financial_rigor.py) tool for financial metrics.
- The **team-lead** aggregates outputs into a structured investment dossier, validates it through [`report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/report_audit.py), and cleans up resources with **TeamDelete**.

## Frequently Asked Questions

### What are the specific analytical perspectives assigned to each agent in the AI Berkshire investment team?

The skill assigns four distinct value-investment philosophies: the **Business Analyst** adopts the 段永平 (Duan Yongping) focus on business quality and moats; the **Financial Analyst** applies the 巴菲特 (Buffett) emphasis on financial metrics and intrinsic value; the **Industry Researcher** channels the 芒格 (Munger) perspective on competitive dynamics and industry trends; and the **Risk Assessor** utilizes the 李录 (Li Lu) framework for management evaluation and risk identification.

### How does the skill ensure data accuracy when agents operate in parallel?

Each agent must verify findings against **two independent data sources** (such as macrotrends and stockanalysis for US equities) and invoke the [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) script for market-cap and valuation cross-checks. This shared data-quality contract ensures consistency across all parallel research streams despite their simultaneous execution.

### Can the /investment-team skill handle private companies or limited data scenarios?

Yes. The skill implements an **AI researchability rating** (A-B-C) that evaluates information richness before team creation. For private companies or data-scarce targets, this rating signals agents to lower confidence thresholds and adjust research depth, while [`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md) demonstrates a similar team-based pattern adapted for non-public entities.

### What happens if one parallel agent fails or returns incomplete data?

The team-lead monitors progress through a live tracking table and can issue `shutdown_request` signals to individual agents. If an agent fails, the lead can re-dispatch the specific **TaskCreate** call for that perspective without restarting the entire workflow, ensuring robustness in the multi-agent parallel research pipeline.