# Research Framework for Unlisted Companies in AI Berkshire: A Multi-Agent Architecture

> Discover the AI Berkshire research framework for unlisted companies. This multi-agent architecture orchestrates specialized agents for validated investment reports from fragmented data.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: architecture
- Published: 2026-07-06

---

**AI Berkshire provides a dedicated multi-Agent workflow for researching private (unlisted) companies, orchestrating six specialized agents through a team-lead coordinator to produce validated investment reports from fragmented data.**

The `xbtlin/ai-berkshire` repository implements a rigorous research framework for unlisted companies within the [`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md) skill. This framework addresses the fundamental challenge of analyzing private companies where public financial statements are unavailable and information is inherently sparse. It transforms fragmented alternative data into a disciplined, transparent investment research deliverable through structured parallel processing and strict validation protocols.

## Core Architecture of the Private Company Research Framework

### Team-Lead Coordination Layer

A **team-lead Agent** orchestrates the entire research process, stitching together findings from specialized agents and producing the final consolidated report. This central coordination layer handles cross-validation, arbitrates conflicts between data sources, and ensures the final output maintains methodological consistency across all research dimensions.

### Six Specialized Research Agents

The framework deploys six concurrent agents, each targeting a specific research dimension as defined in [`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md):

- **business-decoder**: Analyzes business models, product matrices, and user ecosystems
- **financial-detective**: Handles financial data reconstruction and valuation modeling
- **competitive-mapper**: Maps industry-wide competitive landscapes and market positioning
- **risk-governance-analyst**: Profiles management quality, governance structures, and risk factors
- **tech-ip-analyst**: Evaluates technology stacks, patent portfolios, and R&D capabilities
- **signal-miner**: Extracts alternative data including recruitment metrics, app analytics, media sentiment, and legal filings

## The Seven-Step Execution Workflow

The research framework for unlisted companies executes through a structured pipeline:

1. **Team Overview**: The lead presents the agent roster and confirms the target company with the user.

2. **Team Creation**: The system creates a named team (e.g., `spaceX-private-research`) to isolate the research context.

3. **Task Definition**: Six concrete tasks are generated, each with a *subject*, *description*, and *activeForm* driving the respective agent (e.g., "拆解SpaceX商业模式、产品矩阵与用户生态").

4. **Parallel Agent Launch**: All six agents launch **concurrently** within a single request, each receiving role-specific prompt templates.

5. **Progress Reporting**: The lead streams updates on agent completion (e.g., "✅ business-decoder completed – 12 key findings") and summarizes interim takeaways.

6. **Cross-Validation & Information Stitching**: The lead cross-checks data points across agents, arbitrates conflicts, and builds a unified "information puzzle" distinguishing high-confidence (🟢), medium-confidence (🟡), and low-confidence (🔴) data.

7. **Final Consolidation**: Outputs a structured report containing a one-sentence conclusion, company snapshot, six-dimensional scorecard, validated data table, signal-consistency matrix, valuation ranges, investment thesis (buy/hold/sell), risk matrix, exit-path analysis, and decision-summary page.

## Data Source Hierarchy and Validation Standards

### Three-Tier Confidence Classification

Every data point in the research framework requires explicit source attribution and confidence labeling:

- **🟢 Primary sources**: IPO prospectuses, regulator filings, and parent-company financials
- **🟡 Secondary sources**: Reputable outlets (36Kr, Bloomberg, The Information), analyst reports, and industry research
- **🔴 Tertiary signals**: Employee reviews, patent databases, app store analytics, and domain footprints

### Cross-Validation Protocols

When sources conflict, the framework lists all values and requires the team-lead to justify the chosen figure. The [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) module provides decimal-precision utilities for market-cap verification and multi-source cross-validation, automatically flagging tolerance breaches before data enters the final report.

## Bias Mitigation and Financial Rigor

The framework enforces strict methodological discipline to prevent narrative bias and ensure accuracy:

- **"Leave-Blank" Principle**: Explicitly forbids fabricating numbers. If a dimension lacks sufficient data, the report must state "信息缺失" (data missing) without speculation or extrapolation.
- **Inverse Thinking**: Implements "芒格逆向检验" (Munger's inverse verification) and first-principles questioning to challenge over-optimistic assumptions and avoid confirmation bias.

Financial calculations and validation logic delegate to **[`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)**, which guarantees decimal-precision and validates cross-source consistency before final report generation.

## Initiating a Private Company Research Session

Execute the research workflow using the slash command:

```bash
/private-company-research SpaceX

```

This command automatically:

- Creates a team named `spaceX-private-research`
- Fires the six specialized agents in parallel
- Streams progress updates to the console
- Produces a final markdown report saved under [`reports/SpaceX/SpaceX-private-YYYYMMDD.md`](https://github.com/xbtlin/ai-berkshire/blob/main/reports/SpaceX/SpaceX-private-YYYYMMDD.md)

## Summary

- **AI Berkshire's research framework for unlisted companies** employs a multi-Agent architecture with six specialized agents (business-decoder, financial-detective, competitive-mapper, risk-governance-analyst, tech-ip-analyst, and signal-miner) coordinated by a team-lead.
- The **seven-step workflow** runs agents in parallel, then cross-validates findings to produce a unified report with explicit confidence indicators (🟢🟡🔴).
- **Data hierarchy** distinguishes primary, secondary, and tertiary sources, with strict validation via [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) to ensure decimal-precision and flag tolerance breaches.
- The **"Leave-Blank" principle** (信息缺失) prevents speculative fabrication when data is missing, while inverse-thinking checks mitigate cognitive bias.
- All functionality is defined in [`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md) within the `xbtlin/ai-berkshire` repository.

## Frequently Asked Questions

### What makes AI Berkshire's approach to unlisted companies unique?

The framework uses a **multi-Agent parallel processing** system where six specialized agents simultaneously investigate different dimensions of a private company. Unlike single-query research, this architecture enables real-time cross-validation between agents, with a team-lead arbitrating conflicts and stitching together a unified "information puzzle" with explicit confidence labels for every data point.

### How does the framework handle missing financial data?

The framework follows a strict **"Leave-Blank" principle** (信息缺失). When insufficient data exists for a specific metric, the system explicitly marks it as missing rather than extrapolating or fabricating values. This prevents over-optimistic valuations and ensures investors understand exactly where information gaps exist, maintaining analytical integrity.

### What tools ensure accuracy in the final reports?

The **[`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)** module provides decimal-precision utilities for market-cap verification and multi-source cross-validation. It automatically flags tolerance breaches and ensures that financial calculations meet rigorous standards before inclusion in the final report, guaranteeing that valuation ranges and financial reconstructions are mathematically sound.

### Can the framework research any private company?

Yes, the methodology is designed to research any unlisted company by leveraging **alternative data sources** through the signal-miner agent. While the analytical framework remains consistent across targets, the depth of the final report depends on available primary and secondary sources, with the confidence scoring system (🟢🟡🔴) transparently indicating data reliability for each finding.