TradingAgents is an open-source Python framework from Tauric Research that copies the structure of a trading firm with LLM agents. It is based on a research paper (arXiv 2412.20138) by Yijia Xiao, Edward Sun, Di Luo and Wei Wang. It was one of GitHub's fastest-rising projects this week, gaining about 5,500 stars to reach roughly 122,000, with about 23,000 forks.
How it works: built on LangGraph, it runs four analysts in parallel (fundamentals, sentiment, news and technical). Their reports go to a bullish and a bearish researcher who debate them, then to a trader who proposes timing and size, and finally to a risk team and a portfolio manager who approve or reject the order. Approved orders go to a simulated exchange, not a real broker. According to the DEV Community write-up, the point of the adversarial debate is to counter the confirmation bias that single-prompt trading bots often show.
Models and features: the README lists OpenAI, Google Gemini, Anthropic, xAI, DeepSeek, Qwen, GLM, MiniMax, OpenRouter, Mistral, Kimi, Groq, NVIDIA NIM, AWS Bedrock, Azure OpenAI, any OpenAI-compatible endpoint and local models through Ollama. Recent releases added checkpoints to resume a crashed run, backtesting, portfolio-aware runs and single-page HTML reports. The latest TradingAgents release is v0.7.1 (October 2026). In November the same team opened early access to a separate model, Trading-R1.
How to try it: it needs Python 3.11 or later. Clone the repository, then run pip install . (or uv pip install .), or use Docker with docker compose run --rm tradingagents. Then start the interactive CLI with tradingagents, or run it non-interactively, e.g. tradingagents --ticker NVDA --date 2026-09-23 --analysts market,news,fundamentals. You need an API key for your chosen model provider, or a local Ollama model.
Licence and limits: Apache-2.0. The project says it is designed for research and is not financial, investment or trading advice; results depend on the model, temperature, period and data quality, and vary between runs. The paper's backtests cover only three large US tech stocks and, as the DEV article notes, do not model slippage, taxes or market impact. Every run also costs API tokens. The repository's Hacker News search shows no discussion this month.
