TradingAgents: 33K Stars for an Open-Source Framework That Simulates an Entire Trading Firm with LLM Agents

Wall Street runs on teams. A fundamentals analyst digs through earnings reports. A sentiment specialist monitors social media chatter. A technical analyst watches RSI and MACD. A trader synthesizes it all into a position. A risk manager decides whether to let it through.

TradingAgents, built by Tauric Research, takes that exact structure and rebuilds it with LLM-powered agents. Seven specialized roles, one LangGraph-based orchestration layer, and an Apache-2.0 license that lets anyone run what amounts to a simulated trading desk on their laptop. The project has crossed 33,300 GitHub stars and 6,400 forks — numbers that put it among the most popular open-source AI finance projects of the past year.

How Seven Agents Run a Simulated Trading Desk

The framework doesn’t use a single LLM to make trading calls. Instead, it breaks down the decision process into distinct roles that mirror a real firm’s workflow:

The Analyst Team handles data gathering across four specializations. The Fundamentals Analyst evaluates company financials — earnings, balance sheets, intrinsic value calculations. The Sentiment Analyst runs scoring algorithms against social media data and public opinion trends. The News Analyst tracks macroeconomic events, policy changes, and market-moving headlines. The Technical Analyst applies indicator-based pattern detection using signals like MACD, RSI, and moving averages.

The Research Team is where things get interesting. Instead of a single synthesizer, TradingAgents deploys both a Bullish Researcher and a Bearish Researcher. They debate the analyst outputs — one building the case for a position, the other poking holes in it. This adversarial structure is designed to reduce confirmation bias, a problem that plagues both human and AI trading systems.

The Trader takes the debated research and makes the actual call: buy, sell, or hold, along with position sizing and timing.

The Risk Manager and Portfolio Manager form the final checkpoint. They evaluate portfolio exposure, volatility risk, and liquidity concerns before approving or rejecting the trade. No position goes through unchecked.

The entire pipeline runs on LangGraph, which handles the agent communication graph, state management, and the structured handoffs between roles. Each agent produces clear, formatted outputs that feed into the next stage — a design choice that reduces the information loss you’d get from raw free-text chaining.

The Numbers Behind the Hype

Tauric Research published the underlying research as an arXiv paper (2412.20138), now on its seventh revision. The paper reports improvements in cumulative returns, Sharpe ratio, and maximum drawdown compared to baseline single-agent and traditional strategy models. The project’s documentation references up to 30.5% annualized returns in backtesting scenarios — a figure that sounds impressive until you consider the caveats.

And there are significant caveats. The framework’s own documentation states plainly: “designed for research purposes.” Performance varies based on which LLM backbone you choose, the model temperature, data quality, and the specific trading period tested. Backtesting results, as any quant will tell you, don’t account for slippage, transaction costs at scale, or the reality that markets behave differently when real money is on the line.

The community response reflects this tension. On social media, the project went viral with posts describing it as “an AI hedge fund that runs on your laptop.” But more experienced traders and developers pushed back. One recurring criticism: TradingAgents automates known analytical edges — factor models, sentiment pipelines, risk overlays — but doesn’t generate genuinely new alpha. It democratizes access to institutional-style analysis workflows, which is valuable, but it’s not a money printer.

The star growth tells its own story. The project has been climbing GitHub’s trending charts consistently, hitting trendshift.io’s front page in March 2026. It has spawned multiple forks and adaptations: a Chinese-language version, an intraday trading variant, and an integration with Alpaca’s brokerage API for paper trading.

Pick Your LLM: Multi-Provider Support in v0.2.0

The February 2026 release of v0.2.0 was a significant upgrade. The framework now supports six LLM providers through a factory pattern:

OpenAI (GPT-5.x series)
Google (Gemini 3.x)
Anthropic (Claude 4.x)
xAI (Grok 4.x)
OpenRouter (access to multiple models through one API)
Ollama (fully local models, no API costs)

Each provider has its own thinking and reasoning configuration optimized for the framework. The Ollama support is particularly notable — it means you can run the entire multi-agent trading pipeline without sending any data to external APIs. For anyone working with proprietary trading strategies or sensitive financial data, that’s a meaningful feature.

Setup requires Python 3.13 and the appropriate API keys. You can run TradingAgents through its CLI (python -m cli.main) for interactive use, or import TradingAgentsGraph directly into your Python code and call .propagate() to get a trading decision programmatically.

The January 2026 Trading-R1 Technical Report added another layer — though details remain sparse, it signals that Tauric Research is working on reasoning-specialized models specifically tuned for trading analysis.

How TradingAgents Stacks Up Against FinRL, FinGPT, and FinRobot

The AI4Finance Foundation maintains three prominent open-source projects in this space, and each takes a fundamentally different approach:

FinRL uses reinforcement learning. Agents learn trading policies through trial and error in simulated environments, without any language understanding. It scored 53% in a recent agentic AI finance benchmark — consistent in quantitative tasks but weaker on anything requiring contextual reasoning.

FinGPT is an open-source financial LLM focused on sentiment analysis, market forecasting, and financial text understanding. It scored highest in the same benchmark at 79%, excelling particularly at financial statement analysis (86%). But it’s a model, not a multi-agent system — it doesn’t simulate collaborative decision-making.

FinRobot tries to bridge the gap. It’s a platform combining LLMs, reinforcement learning, and quantitative analytics for investment research automation. It scored 74% in benchmarks and is the most integrated option from the AI4Finance ecosystem.

TradingAgents occupies a different niche. It’s not trying to be a general-purpose financial AI platform. Its bet is that the multi-agent debate structure — specifically, the adversarial bull/bear researcher dynamic and the multi-stage approval chain — produces better trading decisions than any single model, regardless of how powerful that model is. Whether that thesis holds up in live markets remains unproven, but the architecture is genuinely novel compared to the alternatives.

There’s also a growing ecosystem of commercial AI trading tools — platforms like Trade Ideas, TrendSpider, and Kavout that offer polished UIs and real-time execution. TradingAgents doesn’t compete with these directly. It’s a framework for developers and researchers who want to build and customize their own systems, not a product for retail traders looking for buy/sell signals.

What the Developer Community Is Actually Building With It

The fork ecosystem reveals how developers are adapting TradingAgents beyond its original scope:

AlpacaTradingAgent connects the framework to Alpaca’s brokerage API, enabling paper trading with real market data. This is the closest anyone has gotten to running TradingAgents against live markets (in simulation mode).

TradingAgentsIntraDay modifies the framework for shorter time horizons, moving from the original daily analysis to intraday signals — a significantly harder problem that requires faster data pipelines.

TradingAgentsCN is a full Chinese-language adaptation, reflecting the substantial interest from Chinese developer communities in AI finance tools.

One limitation that developers have flagged: the framework’s sentiment analysis reads data from local files, assuming you’ve already collected and preprocessed it. There’s no built-in Reddit, Twitter, or news API integration. For researchers running historical backtests, this is fine. For anyone trying to build a real-time system, it means bolting on your own data pipeline.

The broader multi-agent trend in AI is clearly fueling interest. Projects like ByteDance’s DeerFlow and Anthropic’s multi-agent code review system have demonstrated that breaking complex tasks into specialized agent roles can outperform monolithic approaches. TradingAgents applies the same principle to an industry where the stakes — and the potential payoff — are measured in dollars rather than developer productivity.

FAQ

Is TradingAgents free to use?
Yes. The framework is open-source under the Apache-2.0 license, which allows both personal and commercial use. However, you’ll need API keys for your chosen LLM provider (OpenAI, Anthropic, Google, etc.), which carry their own costs. Running multiple agents per trading decision means higher token usage than a single-prompt approach. The Ollama option lets you run entirely local models at zero API cost, though performance depends on your hardware.

Can TradingAgents actually make money in live trading?
The framework explicitly states it’s designed for research purposes and is not financial advice. While backtesting results show promising metrics including reported annualized returns of up to 30.5%, backtests are not predictive of live performance. No independently verified live trading results have been published. Treat it as a research and education tool, not a turnkey trading system.

How does TradingAgents compare to using ChatGPT or Claude directly for trading analysis?
The key difference is specialization and structure. Asking a single LLM for a trading recommendation gives you one perspective with whatever biases the model carries. TradingAgents forces structured analysis through multiple specialized lenses, including an adversarial debate between bullish and bearish researchers and a risk management approval step. The architecture is designed to reduce blind spots, though it comes at the cost of higher latency and token usage per decision.

What programming knowledge do I need to run TradingAgents?
Basic Python proficiency and comfort with the command line. You’ll need Python 3.13 installed, the ability to manage API keys via environment variables, and enough understanding to configure which LLM provider and model you want each agent to use. The CLI mode is straightforward for quick experiments; integrating it into custom pipelines via the Python API requires more experience.

Does TradingAgents work for crypto, forex, or options trading?
The current implementation focuses on stock trading. The analyst agents are tuned for equity-style analysis — earnings, fundamentals, stock-specific technical indicators. Adapting it for crypto or forex would require modifying the data sources and analyst prompts. Some community forks are exploring these directions, but no stable adaptation exists for non-equity markets yet.

Top AI Product

Leave a comment Cancel reply