Multi-Agent Value Investing Research Framework: Roles, Skills, and Disagreement Preservation

1. Why Split an Investment Methodology into Multiple Agents

Traditional intelligent Q&A systems usually have a single analysis entry point. When a user asks whether a company is worth investing in, the system outputs the business model, financial condition, competitive advantage, and valuation conclusion all at once. This seems complete, but it has three clear problems.

First, the analysis process is easily influenced by a single narrative. Once the model judges a company to be excellent, it tends to look for supporting evidence afterward and ignore counterexamples.

Second, responsibilities for different dimensions of research are mixed together. Business model, financial quality, management capability, and risk review should be completed from different perspectives. If all of them are assigned to the same agent, the output easily becomes a long article with a complete structure but convergent views.

Third, the weighted averaging mechanism common in multi-agent systems erases disagreement. Multiple agents give judgments separately, and the system then calculates a composite conclusion. The result looks safe, but it may hide the conflicts that really matter.

Therefore, a design better suited to investment research is to let multiple agents represent different thinking frameworks, complete analysis independently, and preserve the disagreements among them. The end user sees not a vague average answer, but a set of judgments that can be verified, challenged, and tracked.

2. Responsibilities of Four Investment Agents

The framework can abstract four investment philosophies into four independent research agents. The focus here is not to replicate personal views, but to convert methodologies into stable, executable analysis processes.

2.1 Long-Term Value Agent: Focus on Intrinsic Value

The Buffett-style agent mainly answers the following questions:

• Does the company have a clear and sustainable profit model?
• Do revenue and profit come from real operations rather than one-off factors?
• Can the company continuously generate free cash flow over a long period?
• Can management allocate capital reasonably?
• Is there sufficient margin of safety between the current price and intrinsic value?

This agent should focus on the company’s long-term operating quality rather than chasing short-term hype. Its output needs to distinguish facts, inferences, and assumptions, and avoid replacing evidence with adjectives such as “excellent” or “leading.”

2.2 Inversion Agent: Specifically Seek Failure Paths

The core task of the Munger-style agent is not to prove that the company is worth investing in, but to actively seek reasons it may fail. For example:

• What is the company’s most fragile revenue source?
• Which cost or competitive factors may be underestimated by the market?
• If performance deteriorates over the next three years, where is it most likely to start?
• What overly optimistic assumptions are embedded in the current valuation?
• Are there irreversible risks in management, industry structure, or the regulatory environment?

This agent must be “counterevidence first.” It cannot simply repeat the conclusions of other agents. Instead, it should try to construct stress scenarios and point out what information, once it appears, should overturn the existing judgment.

2.3 Business Model Agent: Study the Moat and Competitive Structure

The Duan Yongping-style agent focuses on why the company can continue making money and whether that capability is easy to replicate. Main analysis content includes:

• Does the product or service truly solve high-frequency user needs?
• Do users choose the company because of brand, cost, channel, network effects, or other factors?
• Can competitors weaken the advantage through price cuts, imitation, or technological substitution?
• Does growth depend on continuous investment, or can it form a virtuous cycle?
• Is the business model simple enough for ordinary investors to understand?

This agent should not discuss short-term stock prices too much. It should explain the source, duration, and decay of the company’s competitive advantage. In particular, it needs to distinguish between “high market share” and “a true moat.”

2.4 Concentrated Investment and Management Agent: Focus on Cognitive Boundaries

The Li Lu-style agent can focus on the relationship among enterprise value, management, industry change, and investor cognition. Typical questions include:

• Does the company’s long-term development reasonably match the macro environment?
• Does management have long-term orientation, integrity, and rational decision-making capability?
• Do investors truly understand the company rather than rely on market consensus?
• Are the key variables affecting enterprise value few enough and trackable?
• If only three judgment bases could be retained, which factors matter most?

This agent emphasizes circle of competence and cognitive boundaries. For information that cannot be verified, it should explicitly mark it as unknown rather than use speculative language to fill gaps.

3. System Architecture: From Role Prompts to Composable Skills

A complete framework should not be just four prompts. It should be designed as a reusable skill set. Skills can be divided into four categories according to the research workflow.

3.1 Research Skills

The research module is responsible for collecting and organizing basic information, including:

• Company overview and business breakdown;
• Industry structure and competitive landscape;
• Management and capital allocation record;
• Historical operating changes;
• Key risks and controversial matters.

Research output should build a fact list and attach source, time, and credibility to each important conclusion. For information with an unknown source, it should be placed separately in “items to verify” and cannot directly enter the final conclusion.

3.2 Financial Skills

The financial module analyzes revenue, profit, cash flow, balance sheet, and return on capital. In addition to single-year data, it should focus on trends and quality, for example:

• Is revenue growth accompanied by abnormal expansion of receivables?
• Does profit growth depend on non-recurring gains?
• Has operating cash flow lagged behind profit for a long time?
• Does capital investment bring corresponding returns?
• Could the debt structure amplify operating risk?

Financial analysis should not only provide metric tables; it should also explain the relationships among metrics. For example, profit growth while cash flow deteriorates may indicate changes in collections, inventory, or the business model.

3.3 Screening Skills

The screening module is used to judge in a relatively short time whether a research object is worth further study. Industry funnels, quality screening, and investment checklists can be designed.

Among them, the investment checklist should use clear binary conclusions: pass or fail. Do not output vague results such as “basically pass,” “partially meets,” or “cautiously observe”; otherwise, the screening tool loses its decision value.

Checklist content can include:

1. Can the business model be clearly explained?
2. Is the core competitive advantage supported by evidence?
3. Is financial quality stable?
4. Is management credible and rational?
5. Does valuation provide a margin of safety?
6. Are the major failure scenarios acceptable?

If key questions cannot be verified, the rule should judge them as fail, rather than pass with a large number of disclaimers.

3.4 Portfolio Skills

The portfolio module focuses on the impact of a single enterprise after it enters the overall portfolio, including:

• Does the holding logic still hold?
• Have the company’s fundamentals changed?
• Does a valuation change alter the risk-return ratio?
• Is there hidden concentration across assets?
• Which news is worth tracking, and which is just short-term noise?

The key to portfolio management is not to predict prices every day, but to maintain the investment thesis. Each holding should store “buy rationale, key assumptions, verification indicators, invalidation conditions, and next review time.”

4. Parallel Collaboration and Disagreement Preservation

The core of a multi-agent system is not “more agents,” but clear boundaries between agents. A reasonable execution flow can be divided into five steps.

Step 1: Unified Input

All agents receive the same basic materials, including company profile, financial data, industry information, and user question. This reduces false disagreement caused by input differences.

Step 2: Independent Analysis

The four agents run separately and are forbidden to read other agents’ intermediate conclusions. Each agent must independently complete:

• Core judgment;
• Supporting evidence;
• Biggest risk;
• Unknown information;
• Confidence in the conclusion.

Step 3: Conflict Extraction

The aggregation agent is not responsible for “reconciling everything,” but for identifying conflicts. For example:

• One agent believes the moat is solid, while another believes the competitive advantage is easy to replicate;
• One agent believes valuation is reasonable, while another believes growth expectations are too high;
• One agent believes management’s capital allocation is excellent, while another points out clear problems in historical decisions.

Conflicts must be presented as questions and annotate the facts on which the disagreement is based.

Step 4: Evidence Review

Create review tasks for each key disagreement. For example, for the judgment “cash flow quality is declining,” further check the collection cycle, inventory changes, and one-off items; for the judgment “a moat exists,” compare customer retention, competitor entry costs, and price changes.

Step 5: Layered Conclusion Output

The final report is recommended to be divided into three layers:

• Consensus facts: content accepted by multiple agents and supported by data;
• Key disagreements: conflicting judgments reached by different agents based on different assumptions;
• Decision conditions: what future information should raise, lower, or revoke the current judgment.

This result has more research value than a single “buy or not buy” because it tells users what to observe next.

5. Hard Constraints in Prompt Design

To avoid vague agent output, prompts need to include explicit rules.

5.1 No Evidence-Free Adjectives

Words such as “excellent,” “robust,” and “leading” must be paired with factual basis. If no evidence can be provided, rewrite as “currently insufficient information.”

5.2 Force Separation of Facts and Inferences

The following format is recommended:

Facts:
- Disclosed or verifiable information
Inferences:
- Judgments formed based on facts
Assumptions:
- Premises requiring further verification
Invalidation conditions:
- Changes that would overturn the current conclusion

5.3 Force Counterexamples

Each agent must provide at least one piece of evidence supporting the current view and one piece of evidence that could overturn it. This reduces the risk of one-sided argumentation.

5.4 Do Not Replace Judgment with an Average Score

The system can count the distribution of conclusions across agents, but cannot directly convert it into a “composite score.” A score can only be auxiliary information and cannot obscure material disagreements.

6. Practical Usage and Engineering Implementation

In engineering implementation, each skill can be saved as an independent text configuration file and invoked through a unified command entry point. Typical usage can be abstracted as:

/investment-team <company>
/financial-review <company>
/industry-screen <industry>
/investment-check <company>
/portfolio-review <current-portfolio>
/information-pulse <company>

To ensure maintainability, skill files should include the following parts:

• Task objective;
• Input format;
• Execution steps;
• Output template;
• Prohibited actions;
• Failure handling;
• Questions to hand off to other skills.

If deployed in an enterprise environment, data reading, document parsing, structured storage, logging, and human approval can be designed as independent modules. When sensitive data is involved, permission control, de-identification, and audit records should be added.

7. Limitations and Usage Boundaries

This type of system is essentially a research assistance tool, not an automatic investment decision maker. It may still face the following problems:

• Data sources are incomplete or lagging;
• Agents may misinterpret financial definitions;
• Historical methodologies may not apply to all industries;
• Multiple agents may share the same erroneous information;
• Complex business models are difficult to simplify into fixed templates;
• Valuation conclusions depend heavily on assumptions rather than pure calculation results.

Therefore, system output should be treated as a “reviewable research draft.” Truly important judgments still require human review, especially data authenticity, accounting definitions, legal risks, and valuation assumptions.

Conclusion

Designing different investment methodologies as multiple agents with clear responsibilities is valuable not because it simulates the tone of famous investors, but because it converts long-term value, inversion, business models, and cognitive boundaries into repeatable research workflows.

A mature intelligent investment research framework should achieve three things: let agents divide work rather than duplicate labor, let evidence support judgments rather than pile up opinions, and let disagreements be recorded and tracked rather than averaged away. Only then will multi-agent technology move beyond the demonstration level and truly become an engineering tool for researchers to discover problems, validate assumptions, and continuously review their work.