How the Earnings-Team Skill Orchestrates Four Agents for Comprehensive Earnings Analysis

The /earnings-team skill in the xbtlin/ai-berkshire repository coordinates four master-perspective agents—段永平, 巴菲特, 芒格, and 李录—plus supporting Editor and Reader agents to transform raw earnings data into a publish-ready WeChat article through parallel research, synthesis, and editorial refinement.

The ai-berkshire project implements a team-layer workflow that demonstrates advanced multi-agent orchestration for financial analysis. This system leverages the earnings-team skill to deploy four specialized master agents that analyze earnings data from distinct value-investing perspectives, then synthesizes their outputs into rigorous, publication-grade research. The entire pipeline is defined in skills/earnings-team.md and its generated counterpart codex-skills/earnings-team/SKILL.md.

Architecture of the Four-Agent Research Pipeline

The earnings-team skill structures its workflow into three distinct phases, each managed by the Team Lead coordinator.

Phase 1: Parallel Master Analysis

During the research phase, the Team Lead launches four master agents in parallel within a single message, as specified in the "启动4个并行研究Agent" section of the skill definition. Each agent receives the same raw materials—earnings transcripts, call transcripts, shareholder letters, and prior period data retrieved from IR pages, SEC EDGAR, HKEX, and 巨潮—along with a material-availability rating (A/B/C) that determines analysis depth.

• Agent 1 (段永平) evaluates business fundamentals and "生意本质" (essence of the business).
• Agent 2 (巴菲特) audits financial quality and cash-flow characteristics.
• Agent 3 (芒格) analyzes competitive dynamics and moat sustainability.
• Agent 4 (李录) identifies risk signals and downside scenarios.

The Team Lead continuously updates a progress board (跟踪进度) so users can monitor which masters are still processing.

Phase 2: Team Lead Synthesis

Once all four agents complete their independent analyses, the Team Lead aggregates their reports following the "Team Lead 合成研究报告" guidelines. This synthesis step identifies consensus views, contradictions, and gaps in coverage, ensuring cross-validation of data while highlighting substantive disagreements between the masters. The output is a unified research report draft that serves as the foundation for publication.

Phase 3: Editorial Processing and Review

The publication phase introduces two additional agents:

• Agent 5 (Editor) rewrites the research draft into a concise, reader-friendly WeChat article format.
• Agent 6 (Reader) evaluates the draft for readability, information value, credibility, and actionability.

The Team Lead incorporates feedback from both agents, refines the article accordingly, and executes a final compliance check before storing all artefacts under reports/{company_name}/.

Orchestration Mechanics

Parallel Execution and Data Validation

The skill spawns the four master agents simultaneously using a single command, allowing each to perform deep, independent searches without sequential blocking. This parallelism is managed through the underlying Agent Tool that handles background material gathering (获取一手资料) before agent activation.

Financial rigor is maintained through specialized utilities. Both the 巴菲特 Agent and the final audit step invoke financial_rigor.py to cross-validate revenue, cash-flow ratios, and valuation metrics. The Team Lead subsequently runs report_audit.py to execute a compliance extract-and-verdict check, certifying the article before release.

Implementation Examples

Invoking the Skill

To initiate the four-agent analysis workflow, use the slash command with the company name and quarter:

/earnings-team 腾讯 2025Q4

This command prints a progress board, spawns the four master agents, generates the consolidated research draft, and produces the edited article with reviewer feedback.

Output Artefact Structure

The Team Lead automatically organizes all generated files under a dedicated directory:

reports/腾讯/
├── 腾讯-earnings-2025Q4.md               ← Final WeChat article
├── 腾讯-earnings-2025Q4-研究底稿.md       ← Four-masters synthesis report
├── 腾讯-earnings-2025Q4-段永平.md         ← Business fundamentals analysis
├── 腾讯-earnings-2025Q4-巴菲特.md         ← Financial quality audit
├── 腾讯-earnings-2025Q4-芒格.md           ← Competitive dynamics analysis
├── 腾讯-earnings-2025Q4-李录.md           ← Risk signals analysis
└── 腾讯-earnings-2025Q4-读者评审.md       ← Reader evaluation report

Tool Integration Within Agents

Individual agents embed tool commands to guarantee data integrity. For example, the 巴菲特 Agent might invoke:

python3 tools/financial_rigor.py cross-validate \
  --metric "revenue" --values 12345 12400 \
  --sources "SEC" "IR"

This command validates revenue figures against multiple sources before inclusion in the final analysis.

Key Files in the Repository

• skills/earnings-team.md — Canonical skill definition outlining all agents, phases, and orchestration logic.
• codex-skills/earnings-team/SKILL.md — Generated Codex-compatible version synchronized by scripts/sync-codex-skills.py.
• tools/financial_rigor.py — Cross-validation utilities for revenue, cash-flow, and valuation metrics used by the 巴菲特 Agent.
• tools/report_audit.py — Final audit execution for extract-and-verdict compliance checks.
• README.md — High-level documentation of the team-layer concept and agent architecture.

Summary

• Parallel Analysis: The earnings-team skill simultaneously activates four master agents (段永平, 巴菲特, 芒格, 李录) to analyze earnings from distinct value-investing perspectives.
• Synthesis and Conflict Resolution: The Team Lead aggregates independent reports, identifies contradictions, and produces a unified research draft that maintains analytical rigor.
• Editorial Workflow: Dedicated Editor and Reader agents refine content for publication, ensuring the final output is both credible and accessible.
• Automated Validation: Integration with financial_rigor.py and report_audit.py ensures data integrity and compliance before article release.
• Structured Output: All artefacts are organized under reports/{company_name}/ with clear separation between raw master analyses, synthesis drafts, and final publications.

Frequently Asked Questions

How does the earnings-team skill resolve conflicts between the four master agents?

The Team Lead explicitly identifies contradictions and consensus areas during the synthesis phase, cross-referencing claims against primary source data from the "获取一手资料" step. This produces a balanced research draft that acknowledges divergent viewpoints while prioritizing factually validated conclusions.

What external data sources do the agents query during research?

Each master agent independently fetches primary filings from IR pages, SEC EDGAR, Hong Kong Exchange (HKEX), and 巨潮. This multi-source approach ensures comprehensive coverage regardless of whether the company is US-listed, HK-listed, or A-share.

How does the system ensure financial data accuracy?

The 巴菲特 Agent and final audit steps invoke financial_rigor.py to cross-validate metrics like revenue and cash-flow ratios against multiple sources. Additionally, report_audit.py performs final compliance verification before granting publication approval.

Can the earnings-team skill analyze companies without recent earnings reports?

The skill assigns a material-availability rating (A/B/C) during the initial data gathering phase. This rating automatically adjusts the analysis scope and depth expectations when primary filings or transcripts are limited, allowing the workflow to proceed with reduced but rigorous coverage.

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