Multi-Agent vs Single-Agent: Why One Chatbot Isn't Enough

You open ChatGPT and type "Analyze HPG stock for me." You receive a 500-word passage, and after reading it, you still do not know whether to buy or sell. No specific numbers, no industry comparison, no clue where the data came from.

The issue is not that ChatGPT is bad — it is that the Single-Agent architecture has inherent limitations when handling complex tasks like stock analysis.

3 Major Limitations of a Single Chatbot (Single-Agent)

1. Hallucination — Confidently Wrong

A single-agent LLM generates text based on probability. When it lacks accurate data, it still produces an answer — and does so with remarkable confidence. You ask for HPG's P/E ratio, and it may return a plausible-sounding number that is completely wrong because its training data is 6 months old.

In investing, wrong data = wrong decisions = lost money. This is not a minor issue.

2. No Real-Time Data

ChatGPT, Claude, Gemini — all have knowledge cutoffs. Vietnamese market data is even scarcer since most training data is in English. The result: a chatbot knows more about Apple than Vinamilk, and more about the S&P 500 than the VN-Index.

3. Not Deep Enough

A general-purpose model can write poetry, code in Python, translate text, and "analyze stocks." But serious stock analysis requires: standardized financial data, structured analysis frameworks, peer comparisons, and technical analysis. A model that does everything does nothing deeply.

Multi-Agent: One Agent, One Job

The Multi-Agent architecture solves all three problems by dividing work among specialized agents.

Think of it this way: a Single-Agent is like a freelancer — knowledgeable about many things but only moderately good at each. A Multi-Agent system is like a team of experts — each person excels at exactly one thing, coordinated by a team lead.

How Multi-Agent Works in FinStock

The Coordinator acts as team lead — receives your request, understands context, and assigns it to the right agent. It does not perform the analysis itself but ensures the right agent gets the right task.

The Planner develops the strategy — deciding what data is needed, which analysis framework to use, which companies to compare against, and what output format to produce. Each question gets a different plan.

The Researcher is the data specialist — querying directly from a database of ~7 million records. Financial reports, historical prices, insider transactions, foreign investor flows, news. This agent does not generate data — it queries real data.

The Reporter is the analyst who writes the report — receiving data from the Researcher, synthesizing it into a structured analysis with comparisons, conclusions, and action items. Output can be text, PDF, or PPT.

Visual Comparison: Single vs Multi-Agent

Same Question, Two Very Different Results

Let's see what happens when both receive: "Analyze HPG."

Single-Agent Response

A text block of about 300-500 words. Content typically includes: an introduction saying HPG is Vietnam's largest steel company, some generic observations about the steel industry, and advice like "you should carefully consider before investing." No specific numbers from the latest quarter, no comparison with NKG or HSG, no technical analysis.

After reading, what have you learned? Almost nothing.

Multi-Agent Response

A clearly structured report:

• Overview: HPG's position in the industry, market share, trends
• Financials: Revenue, profit, P/E, ROE, D/E — figures from the latest quarter
• Industry comparison: HPG vs NKG vs HSG vs POM — comparison across 106 criteria
• Technical: Price trends, volume, support/resistance, signals
• News: Important recent events affecting HPG
• Conclusion: Overall assessment with specific action items

After reading, you have enough information to make a decision.

Why Doesn't Everyone Use Multi-Agent?

Short answer: it is hard to build.

Multi-Agent systems are many times more complex than single-agent ones. You need to design: how agents communicate, how to handle agent failures, how to maintain speed with parallel execution, and how to manage data pipelines for domain-specific data.

For the Vietnamese stock market, the challenge is even greater: data is fragmented (CafeF, Vietstock, SSI, HOSE...), formats are inconsistent, and official APIs are scarce. FinStock built FinData — a dataset of ~7 million records standardized across 106 criteria — as the foundation for the Researcher Agent.

When Is Single-Agent Still Fine?

To be fair, a Single-Agent is still useful for:

• General knowledge Q&A: "What is P/E?", "How is ROE calculated?"
• Brainstorming initial investment ideas
• Explaining financial concepts

But when you need analysis based on real data, with a structured process and real depth — that is where Multi-Agent excels. And in investing, accuracy and depth of analysis directly affect your portfolio.