I Built an AI Team That Researches Stocks and Debates Investment Decisions 🤖

What if you had an entire team of AI agents researching stocks for you?

Imagine one agent analyzing stock charts, another digging through financial reports, and another tracking the latest market news.

Now imagine two more AI agents arguing about whether you should buy or avoid a particular stock.

And instead of simply combining their opinions, the system makes them challenge each other’s conclusions before arriving at a final investment recommendation.

Sounds like something a hedge fund might build, right?

Well, there’s an open-source project that does exactly this.

It’s called TradingAgents, developed by I Built an AI Team That Researches Stocks and Debates Investment Decisions 🤖, and it uses a multi-agent AI architecture to simulate how a financial research team collaborates.

I recently explored this project, studied its architecture, set it up, and put the agents to work analyzing one of the most talked-about companies in the stock market: NVIDIA (NVDA).

The result was interesting, but what fascinated me even more was how the agents worked together.

In this article, I’ll explain how TradingAgents works, walk through its architecture, show you how to get started, and share what happened during my experiment.

If you’ve ever researched a stock before investing, you know how overwhelming the process can become.

And here’s something investors sometimes overlook:

A great company isn’t necessarily a great stock to buy at any price.

A business might be growing rapidly and dominating its industry, but if its stock price already reflects unrealistic growth expectations, investing could still be risky.

Now, you could ask ChatGPT or another LLM:

But a single AI response might overlook important factors or lean too heavily toward one interpretation.

Instead of asking one AI model to handle everything, we could assign different research responsibilities to specialized AI agents.

One focuses on financial fundamentals. Another analyzes technical indicators. Another investigates risks.

Then we make them collaborate, debate, and challenge each other’s conclusions.

That’s the fundamental idea behind TradingAgents.

TradingAgents is an open-source multi-agent framework developed by Tauric Research.

It’s designed to simulate the collaborative decision-making process of a financial research and trading organization.

Rather than relying on a single LLM prompt, TradingAgents coordinates multiple specialized agents through a structured workflow.

Each agent has a specific responsibility, and information flows between them as the research progresses.

The interesting part is that the agents don’t simply generate independent reports.

They share findings, debate competing investment theses, evaluate risk, and eventually produce a consolidated recommendation.

The workflow begins with four specialized research agents.

Each examines the company from a different perspective.

The Market Analyst focuses on technical analysis.

For example, the RSI can help identify whether a stock has experienced unusually strong buying or selling momentum.

The analyst uses available technical indicators to understand what’s happening with the stock’s price behavior.

The Fundamental Analyst investigates the underlying business.

It examines financial information such as revenue, earnings, profitability, and other available company data.

What do its fundamentals suggest about its business performance?

Are there financial weaknesses investors should consider?

For a company like NVIDIA, this analysis might involve examining revenue growth, margins, profitability, and the sustainability of demand for its products.

Financial performance alone doesn’t tell the entire story.

A company’s outlook can change significantly because of new developments.

The News Analyst investigates relevant information such as:

For example, changes in semiconductor export restrictions or major AI infrastructure spending announcements could influence how investors evaluate NVIDIA.

The Sentiment Analyst examines available public discussions and sentiment-related information.

Its goal is to understand how investors and the broader market are discussing the company.

This can provide useful context about market expectations, enthusiasm, and concerns.

However, sentiment isn’t necessarily an indicator of whether a stock is fundamentally undervalued or overvalued.

The four analysts can gather information in parallel.

Rather than forcing a single AI agent to perform every research task, the system distributes responsibilities across specialized agents.

Their findings become the foundation for the next stage.

And this is where things get particularly interesting.

Once the analysts complete their research, their findings are passed to two additional agents.

The Bull Researcher and the Bear Researcher.

These agents have opposing responsibilities.

The Bull Researcher builds the strongest possible argument for investing in the company.

For example, a bullish NVIDIA investment thesis might emphasize:

The Bear Researcher takes the opposite position.

Its responsibility is to challenge the bullish argument and identify potential weaknesses.

These agents actually debate each other.

The Bull gets an opportunity to respond.

And the discussion can continue through multiple configurable rounds.

Instead of simply asking an LLM to list pros and cons, TradingAgents creates a structured exchange between competing investment perspectives.

After the debate, a Research Manager evaluates the arguments and produces an investment plan.

This approach doesn’t guarantee better investment decisions, but it provides a mechanism for examining competing interpretations before reaching a conclusion.

After the Research Manager produces an investment plan, the workflow moves to the Trader Agent.

Despite its name, the Trader Agent in my demonstration wasn’t executing any actual trades.

Instead, it translated the research into a more concrete investment proposal.

That proposal then went through another review process.

This time, three risk-focused agents evaluated it.

Examines the investment proposal from a perspective that accepts greater risk in pursuit of potential returns.

Focuses on downside protection, uncertainty, and reasons to avoid excessive exposure.

Attempts to balance potential opportunities against the associated risks.

These perspectives are used to challenge the investment proposal.

Finally, the Portfolio Manager evaluates the research, proposed action, and risk discussion to produce the final recommendation.

So the overall workflow looks something like this:

STOCK / COMPANY | v RESEARCH ANALYSTS | +------------+------------+ | | | Market Fundamental News Analyst Analyst Analyst | | | +------------+------------+ | Sentiment Analyst | v RESEARCH DEBATE | +------+------+ | | Bull Agent Bear Agent | | +------+------+ | v Research Manager | v Trader Agent | v RISK ANALYSIS | +----------+----------+ | | | Aggressive Conservative Neutral | | | +----------+----------+ | v Portfolio Manager | v FINAL RECOMMENDATION Enter fullscreen mode Exit fullscreen mode Simplified conceptual representation of the TradingAgents workflow. The analyst stage is designed to support parallel research.

This is what makes TradingAgents an interesting example of agentic AI and multi-agent orchestration.

It’s a structured workflow in which different agents contribute information, challenge conclusions, and participate in a decision-making process.

Now that we understand the architecture, let’s look at how to run it.

TradingAgents is open source, so you can inspect the implementation, modify the agents, and experiment with different configurations.

git clone https://github.com/TauricResearch/TradingAgents.git cd TradingAgents Enter fullscreen mode Exit fullscreen mode Open the project in your preferred code editor.

I used Python 3.13 and uv for environment management.

uv venv --python 3.13 Enter fullscreen mode Exit fullscreen mode Activate it on macOS or Linux:

source.venv/bin/activate Enter fullscreen mode Exit fullscreen mode This keeps the project’s Python dependencies separate from other applications installed on your machine.

Install the project using the dependency instructions provided in the repository.

For example, when installing from the project directory with uv:

uv pip install -e. Enter fullscreen mode Exit fullscreen mode The exact installation steps may change between releases, so check the repository’s latest README before proceeding.

TradingAgents supports configurable AI model providers.

Depending on the version and configuration, you can use supported providers such as OpenAI, Google, or Anthropic.

The repository includes an example environment configuration file.

cp.env.example.env Enter fullscreen mode Exit fullscreen mode Then configure the credentials required by your selected provider.

In my demonstration, I used the framework’s supported ChatGPT subscription authentication flow rather than manually configuring an API key.

For versions supporting this option, the login command is:

tradingagents login chatgpt Enter fullscreen mode Exit fullscreen mode Follow the authentication instructions displayed in your terminal.

Note: Subscription-based authentication depends on the installed TradingAgents version and its supported integration. A ChatGPT subscription is not generally interchangeable with OpenAI API credits.

Once everything is configured, launch the CLI:

tradingagents Enter fullscreen mode Exit fullscreen mode The interface guides you through several configuration options.

For my experiment, I selected NVIDIA (NVDA) and enabled all four research analysts.

I also chose the shallow research configuration to keep the demonstration relatively quick.

Research depth affects how many debate and discussion rounds the agents perform.

Once configured, the agents begin researching the company.

With everything running, I gave the system NVIDIA as the company to investigate.

The workflow started with the research analysts gathering information.

The Market Analyst examined technical indicators.

The Fundamental Analyst investigated the company’s financial position.

Meanwhile, the News and Sentiment Analysts collected information from their available data sources.

Once the initial research was complete, the system moved through the debate and risk evaluation stages.

Eventually, TradingAgents generated a detailed report containing the findings from the different agents.

The report included technical analysis, financial research, sentiment observations, news analysis, and the final investment thesis.

After completing the analysis, the Portfolio Manager produced its final recommendation:

In portfolio terminology, an underweight position generally means maintaining less exposure to an asset than a relevant benchmark or reference allocation.

But I wouldn’t interpret this result as proof that NVIDIA is a bad investment.

The recommendation reflects the model-generated research, available information, and configuration used during this particular run.

I haven’t established whether this recommendation would outperform a benchmark or lead to better investment outcomes.

Generating a convincing investment thesis is not the same as demonstrating predictive accuracy.

What I found useful was the structured research process.

Instead of receiving a short answer from a chatbot, I could examine the different research perspectives that contributed to the final decision.

That makes TradingAgents particularly interesting as an investment research assistant.

Another feature worth exploring is the framework’s built-in memory and historical evaluation capabilities.

After completing an analysis, TradingAgents can save information about the research decision in a local journal.

Initially, the outcome is marked as pending.

That makes sense because the system doesn’t know what will happen to the stock after the recommendation.

Once sufficient trading sessions have passed, the framework can evaluate the stock’s subsequent performance against a benchmark.

In the configuration I explored, the default evaluation window was five trading days.

Based on the evaluation, an AI model can generate a reflection about the previous research.

Which assumptions turned out to be incorrect?

What lessons might be useful for future research?

Those reflections can then be retrieved as context for subsequent analyses.

This creates a basic feedback mechanism.

However, it’s important to understand that memory-based reflection is not the same as training or fine-tuning the underlying LLM.

The framework can reuse previous lessons as contextual information, but that doesn’t automatically mean its investment predictions improve over time.

Demonstrating that would require systematic evaluation.

After experimenting with TradingAgents, I started thinking about what else we could build using this architecture.

Because the project is open source, there are several interesting directions to explore.

Instead of analyzing one company at a time, imagine providing the system with a watchlist of 50 stocks.

The agents could research each company, evaluate its investment thesis, and produce a shortlist of opportunities worth investigating further.

You could then compare the companies using consistent research criteria.

Another possibility is building a dashboard that tracks the agents’ research over time.

This could make it easier to compare how investment theses evolve as new information becomes available.

Different industries require different types of analysis.

For example, we could introduce specialist agents for:

Semiconductors: Manufacturing capacity, AI infrastructure demand, supply chains, and competitive positioning.

Healthcare: Clinical trials, regulatory approvals, product pipelines, and reimbursement risks.

Technology: Revenue growth, recurring revenue, customer retention, and competitive advantages.

These agents could supplement the general research team with more specialized domain knowledge.

This is probably the most interesting direction from a research perspective.

Single LLM vs. Multi-Agent Research vs. Traditional Financial Screening

We could also run historical evaluations while carefully controlling for look-ahead bias and survivorship bias.

That would help answer a much more meaningful question:

Do multi-agent AI systems actually improve investment research quality, or do they simply generate more elaborate reports?

Expand Vivek Shetye Vivek Shetye Vivek Shetye Follow Software Engineer (7+ yrs) building scalable systems and exploring Agentic AI. Sharing practical tutorials on AI workflows, automation, and multi-agent systems. Location Vancouver, BC Joined Apr 14, 2026 • Nov 3 Dropdown menu Copy link Hide Would you trust a team of AI agents to assist with your investment research? Or do you think multiple LLMs debating each other adds unnecessary complexity?