AI Agents in Stock Analysis: Multi-Agent Architecture Explained

1. The Problem: One Chatbot Is Not Enough

Many investors have tried asking ChatGPT or other popular LLMs about Vietnamese stocks. The results are usually disappointing.

Try a simple experiment: ask ChatGPT "Analyze HPG stock." You will receive a generic article with no current data, no industry comparison, and critically — no buy/sell conclusion based on real numbers.

The reasons are clear:

• No real-time Vietnamese data — ChatGPT is trained on older data and is not connected to the HOSE, HNX, or UPCOM exchanges. When you ask about current prices, volume, or P/E, it can only guess or say "I don't have the latest data."
• Hallucinated numbers — When pressed for figures, LLMs tend to "fabricate" them — known as hallucination. Revenue, profit, and EPS can be completely wrong, with no fast way for investors to verify.
• Not specialized for finance — ChatGPT excels at writing emails, translating, and brainstorming. But serious financial analysis requires: standardized financial data, specific analysis frameworks, peer comparisons, and technical analysis. A general-purpose model is not built for this.

So what is the solution? Instead of trying to make one chatbot good at everything, build a system of specialized agents, each with a defined role — like a real analysis team.

2. Multi-Agent Architecture: 4 Specialized Roles

Multi-Agent is an architecture where multiple independent AI Agents collaborate to complete a complex task. Instead of one model doing everything, each agent is designed to excel at one specific job.

Imagine you are the investment director at a fund. When you need to analyze a stock, you do not do everything yourself — you have a team:

Coordinator — The Team Lead

The Coordinator is the first agent to receive your question. Its role resembles an experienced team lead: reading and understanding the intent, determining what type of analysis is needed, and assigning the work to the right people.

When you ask "Comprehensive analysis of HPG," the Coordinator recognizes this as a request for a full analysis — requiring fundamental, technical, and industry comparison — and initiates the appropriate workflow. If you simply ask "What is HPG's P/E?", the Coordinator knows this is a quick data query and handles it more efficiently.

Planner — The Strategist

The Planner receives the brief from the Coordinator and designs a detailed analysis plan: what data to collect, which industry peers to compare against, which screening strategies to apply, and what the final output format should be.

This is the step most chatbots skip. Without a plan, AI responds haphazardly — sometimes analyzing technicals, sometimes just listing news. The Planner ensures every analysis has a consistent logical structure.

For example, with "Comprehensive analysis of HPG," the Planner designs:

1. Retrieve financial statements for the last 4 quarters and full year
2. Calculate valuation metrics: P/E, P/B, EV/EBITDA
3. Compare with HSG, NKG, TLH (steel sector peers)
4. Run technical analysis: price trends, volume, support/resistance
5. Check insider trading and foreign investor activity
6. Synthesize into a report with conclusions

Researcher — The Data Specialist

The Researcher is the hardest-working agent. After receiving the plan from the Planner, it connects directly to the FinData system to query information — stock prices, financial reports, news, insider transactions, foreign investor data, and dozens of other data types.

The critical difference: the Researcher does not "remember" data from training. It queries directly from a database of approximately 7 million records, ensuring every number is accurate and current. Each data point has a clear source that can be traced back.

The Researcher can run multiple queries in parallel — fetching financial statements while checking news and computing technical indicators — all within seconds. This is something a human analyst would spend hours doing.

Reporter — The Analyst

The Reporter receives all results from the Researcher and transforms them into a structured, readable report with clear conclusions. It does not dump raw data on screen — the Reporter analyzes, compares, and provides assessments.

A typical Reporter output includes:

• Company overview — Industry, market cap, ownership structure
• Fundamental analysis — Revenue, profit, margins, key financial ratios
• Valuation — P/E, P/B vs sector and historical averages
• Technical analysis — Price trends, signals, support/resistance zones
• Risks and opportunities — Key factors to monitor
• Conclusion — Overall assessment with a score and recommendation

Importantly: the Reporter does not only create text. FinStock can export reports as professional PDFs, PPT slides, or podcast summaries — depending on your needs.

3. Flow: From Question to Report in 30-60 Seconds

Let's walk through a real example: you open FinStock and type "Comprehensive analysis of HPG".

Step 1: Coordinator receives and classifies (1-2 seconds)

The Coordinator reads the question, identifies it as a comprehensive analysis request (not a simple price query or sector comparison), confirms HPG is listed on HOSE, and initiates the full analysis workflow.

Step 2: Planner designs the plan (2-3 seconds)

The Planner receives the "comprehensive analysis of HPG" brief and creates a plan with 8-10 data query steps. It also identifies peer companies for comparison — HSG, NKG, TLH — and selects screening strategies appropriate for the steel sector.

Step 3: Researcher queries data (15-30 seconds)

This is the most time-consuming but also the most impressive step. The Researcher executes dozens of parallel queries:

• Retrieving HPG's financial statements: balance sheet, income statement, cash flow
• Computing 106 standardized metrics: ROE, ROA, gross margin, net margin, D/E, and more
• Querying price data: historical prices, volume, volatility
• Running peer comparisons against HSG, NKG, TLH on key metrics
• Checking recent insider transactions and foreign investor flows
• Aggregating HPG-related news from the past 30 days

All from a database of ~7 million records, not from AI "memory."

Step 4: Reporter synthesizes the report (10-15 seconds)

The Reporter receives all the raw data, analyzes relationships between metrics, cross-references with historical and sector data, and writes a complete report. The result is a 2,000-3,000 word analysis with tables, comparisons, and clear conclusions.

Total time: 30-60 seconds. Compared to 75-120 minutes manually — approximately 100 times faster.

4. Comparison: ChatGPT vs Multi-Agent System

To see the difference clearly, here is a direct comparison between ChatGPT (representing typical chatbots) and FinStock's Multi-Agent system:

This is not a comparison of "good AI vs bad AI." ChatGPT excels in many areas — content writing, brainstorming, coding. But Vietnamese stock analysis requires specialized data + structured processes, and that is what Multi-Agent architecture is designed to solve.

For a deeper look at the differences between architecture types, read Multi-Agent vs Single-Agent: Which Architecture Fits?.

5. Data: The Foundation of Every Analysis

A Multi-Agent system, no matter how well designed, is meaningless without quality data. This is where FinStock fundamentally differs from generic AI chatbots.

~7 Million Records, Continuously Updated

The FinData system — the data layer behind FinStock — contains approximately 7 million records including:

• Historical stock prices — All tickers on HOSE, HNX, and UPCOM
• Financial reports — Balance sheets, income statements, cash flow statements, by quarter and year
• Financial news — AI-classified and tagged to relevant stock tickers
• Insider transactions — Who bought, who sold, and how much
• Foreign investor flows — Daily net buy/sell by ticker
• Corporate events — Dividends, secondary offerings, M&A

106 Standardized Criteria

Each stock is evaluated across 106 standardized financial criteria, including:

• Valuation: P/E, P/B, EV/EBITDA, P/S, PEG
• Efficiency: ROE, ROA, ROIC, gross margin, net margin
• Growth: Revenue growth, profit growth, EPS (QoQ, YoY, CAGR)
• Financial health: D/E, current ratio, quick ratio, interest coverage
• Cash flow: FCF yield, operating cash flow margin, capex ratio
• Technical: RSI, MACD, Bollinger Bands, volume trends

Standardization enables fair comparisons across companies — you cannot compare a bank's P/E with a tech company's P/E without industry context.

22 Screening Strategies

The Researcher agent does not just retrieve raw data — it applies 22 pre-built screening strategies, including classic methods like CAN SLIM, Magic Formula, Piotroski F-Score, and strategies optimized for the Vietnamese market.

For more on specific stock screening approaches, read 5 Ways to Use AI Agents for Investing.

How the Researcher Accesses Data

The Researcher agent uses MCP Tools (Model Context Protocol) to connect with FinData. MCP is a protocol that allows AI agents to call specialized tools — similar to how you use apps on your phone. Instead of "remembering" data, the agent actively queries exactly the data needed at the time of analysis.

Combined with RAG (Retrieval-Augmented Generation) for news and research, the system ensures all information has a clear source and can be verified — completely eliminating the hallucination problem for financial data.

6. Demo: Experiencing FinStock Step-by-Step

Let's see how a real investor experiences FinStock.

Step 1: Ask a Question

Open FinStock and type: "Comprehensive analysis of HPG"

No need to know specialist terminology, no menus or complex filters. Ask in natural language, as if messaging an analyst.

Step 2: The System Works (30-60 seconds)

On screen, you see the system in action — the Coordinator has classified the request, the Planner is designing the approach, the Researcher is querying data. FinStock shows real-time progress so you know what the system is doing, rather than waiting blindly.

Step 3: Receive the Report

After approximately 30-60 seconds, you receive a complete report including:

• HPG overview: Hoa Phat Group, steel sector, market cap, ownership structure
• Fundamental analysis: Revenue, profit, margins — with YoY comparisons
• Valuation: Current P/E and P/B vs sector average and historical range
• Industry comparison: HPG vs HSG, NKG, TLH on key metrics
• Technical analysis: Price trends, volume, technical signals
• Risks and opportunities: Key influencing factors
• Conclusion: Overall assessment, score, and recommendation

Step 4: Export the Report

Export results as PDF to send to clients, PPT for an investment meeting, or listen to a podcast summary while commuting.

Proven in Practice

FinStock (formerly StockGPT) has been deployed at VPBankS — one of Vietnam's leading securities firms — serving 13,000 users with over 70,000 queries processed. This is not a demo or prototype — it is a production system used daily by thousands of investors.

To understand what AI Agents are and why they matter, read What Is an AI Agent? A Simple Explanation for Investors.

7. The Future: Where Will AI Agents Go?

The Multi-Agent architecture that FinStock uses is just the beginning. Looking ahead, three major trends will shape how AI supports investing:

Autonomous Portfolio Management

Currently, AI analyzes and suggests — you make the decisions. In the future, AI Agents could manage portions of your portfolio according to strategies you define: automatically rebalancing, cutting losses, and taking profits based on predefined rules. The agent would act within the boundaries you set and report back on every decision.

Real-Time 24/7 Monitoring

Instead of analyzing only when you ask, AI Agents will proactively monitor your portfolio and alert you: "HPG just broke through the 28,000 VND resistance level with 3x normal volume — this matches the breakout strategy you follow." The agent does not wait for you to ask — it proactively notifies you when something significant happens.

Multi-Market Analysis

Multi-Agent is not limited to the Vietnamese market. The same Coordinator-Planner-Researcher-Reporter architecture can extend to other markets — enabling cross-market comparisons: "Compare HPG with Chinese and Indian steel companies" becomes a question answerable within minutes.

Deeper Personalization

AI Agents will learn your investing style over time: Are you a value investor or growth investor? What is your risk tolerance? Which sectors do you focus on? Based on this, every analysis will be tailored to your individual needs — no more one-size-fits-all.

In summary: The Multi-Agent architecture — with 4 roles: Coordinator, Planner, Researcher, and Reporter — addresses exactly the limitations that single chatbots cannot overcome in stock analysis. Combined with ~7 million Vietnamese data records, 106 standardized criteria, and 22 screening strategies, the system transforms 75 minutes of manual research into 30-60 seconds of structured analysis. And this is only the beginning.