What Are the Different Types of Agents Available in AI-Berkshire? A Complete Taxonomy

AI-Berkshire provides a modular agent architecture where specialized roles—such as Author, Editor, Research, Business-Insight, and News-Event agents—are defined in YAML skill files and invoked via a generic Agent tool to execute research, writing, and analysis workflows in parallel.

AI-Berkshire is an open-source framework built around a modular Agent architecture designed for financial research and content creation. Unlike traditional monolithic systems, the different types of agents available in AI-Berkshire are not hardcoded classes but rather dynamic roles defined in skill Markdown files. Each agent is configured with specific persona prompts and responsibilities, allowing the system to spawn parallel processes that collaborate on complex tasks like earnings analysis or article publication.

Content Creation Agents (WeChat Article Workflow)

The content creation workflow in skills/wechat-article.md demonstrates a three-agent collaboration pattern for producing polished articles.

Author Agent

The Author Agent generates the first draft of a document. It receives research inputs and target audience parameters, then produces the initial narrative structure and content.

Editor Agent

The Editor Agent refines language, structure, and style. It operates on the Author Agent's output, applying editorial standards and preparing the draft for publication.

Reader Agent

The Reader Agent reviews the draft from the target-audience perspective. It simulates reader reception and suggests improvements for clarity, engagement, and accessibility before final publication.

Research Agents (Multi-Source Intelligence)

For deep-dive research tasks, AI-Berkshire deploys parallel research agents that gather distinct data streams. These appear in skills/wechat-article.md and skills/private-company-research.md.

Research Agent A

Research Agent A gathers core content such as company backgrounds, founding stories, and fundamental business metrics.

Research Agent B

Research Agent B collects industry context, market positioning, and application scenarios relevant to the target company.

Research Agent C (Optional)

Research Agent C performs comparative analysis, examining competitors and alternative solutions to provide benchmarking context.

Investment Analysis Agents (The "Four Masters" Perspective)

The skills/earnings-team.md skill implements a sophisticated financial analysis workflow using four specialized agents modeled after renowned investors. These agents run in parallel to analyze quarterly earnings reports from distinct analytical angles.

Business-Insight Agent (段永平视角)

The Business-Insight Agent analyzes the "business essence" of a company. It evaluates moat durability, management quality, and long-term value creation potential from a business-owner perspective.

Financial-Quality Agent (巴菲特视角)

The Financial-Quality Agent performs rigorous cash-flow analysis, profit-quality checks, and valuation assessments. It frequently invokes bundled CLI tools such as financial_rigor.py to cross-validate metrics and ensure numerical accuracy.

Competitive-Analysis Agent (芒格视角)

The Competitive-Analysis Agent evaluates industry dynamics, competitive positioning, and long-term trends. It assesses how the company fares against its peer group and broader market forces.

Risk-Signal Agent (李录视角)

The Risk-Signal Agent detects hidden risks in management commentary and forward-looking statements. It scrutinizes disclosure quality, accounting anomalies, and governance red flags that might indicate future problems.

News Intelligence Agents (Real-Time Monitoring)

The skills/news-pulse.md skill defines a four-agent system for rapid news attribution and market intelligence scanning.

Event Agent

The Event Agent scans news feeds for company-specific events that could impact stock prices, such as product launches, executive changes, or M&A activity.

Policy Agent

The Policy Agent monitors regulatory and policy developments affecting the target firm, tracking changes in legislation, industry standards, and government announcements.

Peer Agent

The Peer Agent surveys competitor announcements and industry-wide movements to provide context on relative performance and strategic shifts.

Sentiment Agent

The Sentiment Agent gauges market-wide sentiment by analyzing social media trends, analyst notes, and public discourse surrounding the company or sector.

Orchestration and Workflow Management

Team-Lead Agent

The Team-Lead Agent (often configured as a human or high-level AI supervisor) orchestrates parallel agent execution, aggregates intermediate outputs, and applies confidence weighting to produce final reports. This pattern appears across multiple skills including skills/earnings-team.md and skills/investment-team.md.

Technical Implementation Patterns

All agents are invoked via the generic Agent tool defined in skill YAML workflows. The system automatically allocates separate backend processes for each agent instance.

Parallel Execution Requirement

The repository architecture requires that multiple agents must be launched within the same message to enable true parallelism. For example, when invoking the earnings team, you must call the Agent tool four times in a single message:

# Example: Launching the four-master earnings analysis workflow

- type: Agent
  name: BusinessInsightAgent
  prompt: |
    你是段永平，请从生意本质角度解读财报…
- type: Agent
  name: FinancialQualityAgent
  prompt: |
    你是巴菲特，请审计财务质量…
- type: Agent
  name: CompetitiveAnalysisAgent
  prompt: |
    你是芒格，请分析竞争格局…
- type: Agent
  name: RiskSignalAgent
  prompt: |
    你是李录，请挖掘潜在风险信号…

Agent-Specific Prompts and Data Quality

Each agent receives a dedicated prompt template describing its persona (e.g., "You are a senior editor" or "You are Warren Buffett reviewing cash flow"). Agents also receive a data-availability rating (A/B/C) which directly influences analysis depth, allowing graceful degradation when primary data sources are missing.

Cross-Validation Capabilities

Specialized agents can invoke auxiliary utilities for precise calculations. For instance, the Financial-Quality Agent routinely calls financial_rigor.py to validate financial metrics against raw source data.

Summary

• AI-Berkshire agents are role-based configurations defined in skill Markdown files rather than distinct code classes.
• Content agents (Author, Editor, Reader) handle document creation workflows in skills/wechat-article.md.
• Research agents (A, B, C) gather parallel intelligence streams for comprehensive company analysis.
• Investment agents model four distinct investor perspectives (Business-Insight, Financial-Quality, Competitive-Analysis, Risk-Signal) for deep financial due diligence.
• News agents (Event, Policy, Peer, Sentiment) provide real-time monitoring capabilities in skills/news-pulse.md.
• Orchestration relies on a Team-Lead Agent to aggregate parallel outputs and finalize deliverables.

Frequently Asked Questions

How are AI-Berkshire agents different from traditional AI agents?

Unlike traditional systems where agents are implemented as separate code classes, AI-Berkshire agents are declarative roles defined within YAML skill files. Each agent is a configuration of prompts and parameters passed to the generic Agent tool, allowing rapid specialization without code changes.

Can agents in AI-Berkshire run in parallel?

Yes. The architecture explicitly supports parallel execution by design. You must invoke the Agent tool multiple times within the same message to spawn concurrent backend processes. This pattern is essential for workflows like the earnings-team analysis where four investor-perspective agents analyze data simultaneously.

Where are the agent prompts defined in the repository?

Agent prompts are stored directly in the skill Markdown files within the skills/ directory. For example, skills/earnings-team.md contains the specific persona prompts for the Business-Insight, Financial-Quality, Competitive-Analysis, and Risk-Signal agents.

How does AI-Berkshire handle missing data when running research agents?

The system implements a data-quality signalling pattern where agents receive an availability rating (A/B/C). This rating directly influences how deeply the agent investigates source material, ensuring workflows can complete gracefully even when primary financial data is incomplete or unavailable.

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