Armchair Generals
When a Wall Street firm buys a stock, there’s a whole room of people behind it. Analysts read the financials and the news, researchers argue about up versus down, a trader sizes the position, risk management pours cold water on it, and the boss makes the call.
Now somebody has built that room out of AIs, and it has 120,000 stars on GitHub.
What it is
TradingAgents comes from Tauric Research. The authors of its paper (arXiv 2412.20138) are from UCLA and MIT. It went open source at the end of December 2024 and now has 122,000 stars (121,716 when I looked) and over 23,000 forks. It’s written in Python on top of LangGraph, under the Apache 2.0 license. The latest version, 0.7.1, came out at the end of October, and releases arrive about once a month.
It calls itself a “multi-agent trading framework.” In plain English: one stock, one room full of AIs in a meeting.
The meeting
Each analysis runs through the same process a trading firm would:
- Four analysts: the fundamentals analyst reads the financials, the sentiment analyst trawls StockTwits and Reddit for the retail crowd’s mood, the news analyst follows the news and the macro picture, and the technical analyst watches indicators like MACD and RSI. All four work at once, and each turns in a report.
- Two researchers, one bull and one bear: they use the four reports as ammunition and debate head to head, one shouting up, the other shouting down. A research manager referees and comes out with an investment plan.
- A trader: reads the reports and the debate, then decides when to buy and how much.
- Three risk analysts: one aggressive, one conservative, one neutral, who argue over the trader’s plan for another round.
- A portfolio manager: makes the final call. If it’s approved, the order goes out. To a simulated exchange.
Count them up: one stock, twelve AIs in a meeting. Of everything AIs have learned from us, meetings may be the thing they do most convincingly.
What you get at the end is a rating on a five-step scale (Buy, Overweight, Hold, Underweight, Sell) plus the full chain of reasoning: who said what and why, saved as a single web page.
It also keeps a little notebook. Every call it makes gets written down. The next time it runs, it first pulls out the old calls whose time is up and checks them against reality: money made or lost, market beaten or not. Then it writes a paragraph of self-reflection. Before the portfolio manager decides, it flips through the notebook. Learning from your mistakes: the AIs are into that too.
Where the data comes from
- Prices: from Yahoo Finance. US stocks, Hong Kong, China’s A-shares (Kweichow Moutai is
600519.SS), Tokyo, London, India, and Bitcoin while it’s at it. - Financial statements: for US companies, from SEC EDGAR, exactly as originally filed. Apple’s total assets for 2008 were filed as $39.6 billion and restated to $36.2 billion in 2010. Set the analysis date in between, and $39.6 billion is what it reads.
- Macro: the Federal Reserve’s FRED data.
- Sentiment and news: StockTwits, Reddit and news feeds.
In a backtest, the AIs only see what had been published before the analysis date, so nobody peeks at the future. Much of the work in the last several releases went into exactly that, and 0.7.1 fixed a bug that let the model read filings submitted after the analysis date. If you could peek at the future, you wouldn’t need a twelve-person meeting.
Pick whatever models you like: OpenAI, Claude, Gemini, Grok, DeepSeek, Alibaba’s Qwen, Zhipu’s GLM, MiniMax, Kimi, local models through Ollama, or your own ChatGPT Plus or Pro subscription. The analysts and researchers run on a fast model and the two managers on a deep-thinking one, and the two groups can even come from different providers.
The paper’s scorecard
The paper backtested from January 1 to March 29, 2024, and its results table lists three stocks:
| Cumulative return | TradingAgents | Buy and hold |
|---|---|---|
| Apple | 26.62% | -5.23% |
| 24.36% | 7.78% | |
| Amazon | 23.21% | 17.1% |
It beat buy-and-hold on all three, and it beat the best of the old-school strategies like MACD and moving averages. Its Sharpe ratios ran from 5.6 to 8.2, and its maximum drawdown stayed under 2.2% every time.
For a sense of scale on a Sharpe ratio of 8: one study found that Warren Buffett’s Berkshire Hathaway, from 1976 to 2011, had a Sharpe ratio of 0.76.
But that scorecard comes with an asterisk: three months, three stocks, and the team grading its own exam. The README says so itself: backtest results aren’t guaranteed to match any published figure, so treat the framework as scaffolding for research, not a strategy with a fixed return you can copy.
How to run it
You need Python 3.11 or later:
git clone https://github.com/TauricResearch/TradingAgents.git
cd TradingAgents
pip install .
tradingagents
That opens a command-line interface. Choose a ticker, an analysis date, a model provider and how deep to research, then watch the AIs turn in their reports one by one. To skip the menus, pass flags: tradingagents --ticker NVDA --date 2026-09-23. It also runs in Docker, or from Python.
To see whether it holds up over time, there’s a backtest:
tradingagents backtest NVDA,AAPL \
--start 2026-06-01 --end 2026-08-01 --every 7
It runs an analysis every 7 days, and once each holding period is up, scores the call against the market benchmark. You can also tell it what you actually hold, and the trader, risk team and portfolio manager will work from your real positions.
The fine print
- All talk, no trades: orders the portfolio manager approves go to a simulated exchange. There’s no broker integration anywhere in the repo, and it will never place an order in your account.
- They say so themselves: it’s for research and is not investment or trading advice.
- Two runs, two answers: the same stock on the same date can come out differently on a second run. The models are random, and the news and the posts change every day.
- A room full of LLMs is not a quant shop: all twelve are language models. They read reports, write reports and argue, and what wins is the argument that sounds right, not a mathematical model fitted to prices.
- No historical financials for Hong Kong or China A-shares: statements for non-US companies come from Yahoo, which doesn’t date when they were published. To avoid peeking at the future, a backtest just doesn’t get them.
- It still costs money: you pay for the model APIs, and if you use a ChatGPT subscription, the runs count against your plan.
A similar project
ai-hedge-fund, 64,000 stars, MIT license, is another meeting of AI investors, with paper trading and backtesting. The first line of its README says it’s for learning only, and the system doesn’t actually make any trades.
One last thing
For the record: this post is about a tool. It is not a stock tip.
In Chinese, the idiom for an armchair general is “discussing war on paper,” and it comes from Zhao Kuo, a general in the third century BC who knew the military classics by heart. He was all theory until he led 400,000 men into the battle of Changping, and it ended in catastrophe.
The thing is, war on paper was never the problem. The problem was Zhao Kuo taking it to the battlefield. TradingAgents sensibly sticks to paper, which already makes it smarter than Zhao Kuo. If you take its call and place a real order, the Zhao Kuo is you.