TradingAgents, the multi-agent LLM trading framework from Tauric Research, is trending on GitHub again. The repository sits at about 114,000 stars, gained roughly 309 in a single day this week, and recently shipped v0.7.0. The project is worth understanding for one reason beyond finance: it is one of the most popular open examples of a role-playing multi-agent graph, where specialist LLM agents argue before a manager decides.
This guide explains what the TauricResearch/TradingAgents repository does, how to run it, what the paper claims, and what the skeptics on Hacker News say. It is an explainer about an agent architecture, not investment advice, and nothing here recommends trading with it.
TL;DR: the questions people ask first
| Question | Short answer |
|---|---|
| What is it? | A LangGraph framework that simulates a trading firm with LLM agents |
| Who built it? | Tauric Research; the paper is by Yijia Xiao, Edward Sun, Di Luo and Wei Wang |
| Latest release? | v0.7.0 (October 2026) per the README news |
| Is it free? | The code is open source; you pay for model API calls, or use your own ChatGPT plan |
| Does it trade real money? | The README says approved orders go to a simulated exchange |
| Does it work? | Authors report gains in the paper; independent evidence is thin |
| Can I run it locally? | Yes, with Ollama or any OpenAI-compatible server |
How the TradingAgents design works
The README describes a pipeline that copies how a real trading desk is organized. Each role is an LLM agent with its own prompt and tools:
| Stage | Agents | What they do |
|---|---|---|
| Analyst team | Fundamentals, Sentiment, News, Technical | Read filings, social posts, headlines and indicators such as MACD and RSI. They run in parallel |
| Researcher team | Bullish and bearish researchers | Debate the analysts' findings in structured rounds |
| Trader | Trader agent | Turns the reports into a decision with timing and size |
| Risk and portfolio | Risk management, Portfolio Manager | Assess volatility and liquidity, then approve or reject. Approved orders go to a simulated exchange |
The idea is adversarial reasoning. Instead of one prompt producing a verdict, a bull and a bear are forced to argue, and a risk reviewer stress-tests the result. That pattern is familiar from other agent designs covered on explainx.ai, such as multi-agent orchestration patterns and the failure modes described in multi-agent error propagation.
Two model tiers
The Python example in the README sets a "deep think" model for complex reasoning and a "quick think" model for fast tasks. The quick model serves analysts, researchers, debaters and the trader, while the deep model serves the research and portfolio managers. Each tier can run on a different provider, for example the managers on Claude and the rest on OpenAI. The default models named in the README are GPT-6 Sol and GPT-6 Luna, which explainx.ai covered when OpenAI announced GPT-6 Luna at DevDay.
What changed in v0.7.0 and v0.6.0
The repo's news section lists the recent releases:
- v0.7.0 (October 2026): runs on your own ChatGPT Plus or Pro plan, backtests scored from the first close after each decision, an update notice at start-up, and a wheel on each GitHub release.
- v0.6.0 (October 2026): reports saved as a single HTML page, a provider per model tier, past decisions settled for every ticker while analysts work, and company news from Yahoo search while Yahoo's news feed is down.
- v0.5.2 (September 2026): parallel analysts, a CLI that runs without prompts through flags such as
--tickerand--date, and backtests that see only data published by each analysis date. - v0.5.0 (September 2026): point-in-time integrity across dated paths, SEC EDGAR fundamentals served as filed, and backtesting over a ticker and date grid.
The theme across releases is look-ahead bias. Several entries are fixes so that a run dated in the past only sees data that existed then. The README gives an example: Apple's 2008 total assets were filed as $39.6B and later restated to $36.2B, so a run dated in between should read $39.6B. That attention to point-in-time data is the right instinct for any backtest.
How to run it
The README's install path is short. You need Python 3.11 or newer.
git clone https://github.com/TauricResearch/TradingAgents.git
cd TradingAgents
python -m venv .venv
source .venv/bin/activate
pip install .
cp .env.example .env
tradingagents
For unattended runs, the README shows flags that skip the interactive questions:
tradingagents --ticker NVDA --date 2026-09-23 --analysts market,news,fundamentals --save --no-show
From Python, you import TradingAgentsGraph and call propagate with a ticker and a date. A Docker image and a docker-compose profile for Ollama are included. For local models, set llm_provider to ollama or openai_compatible and point it at your server.
Using a ChatGPT plan instead of an API key
New in v0.7.0: choose "Subscription" in the provider menu, or set llm_provider: "chatgpt", then sign in with tradingagents login chatgpt. The README says the sign-in is stored in ~/.tradingagents/chatgpt-auth.json, that runs count toward your plan's usage limits, and that the plan does not accept temperature or max tokens. If you try that, read the terms of your plan first and expect limits to apply; a multi-agent debate makes many model calls per ticker.
Memory and reflection
The README says a memory log is always on. Each completed run appends its decision to a local markdown file. On later runs, the system settles earlier decisions whose holding period has passed by fetching realised return, both raw and alpha against a regional benchmark, and writes a short reflection. The Portfolio Manager then reads recent decisions on the same ticker plus lessons from others. That is a simple form of self-evaluation, and it connects to the broader question of how agents learn from outcomes, discussed in explainx.ai's guide to agent skills.
What the paper claims
The arXiv abstract describes the framework as simulating a collaborative trading environment and says experiments show "notable improvements in cumulative returns, Sharpe ratio, and maximum drawdown" over baseline models. The paper was first submitted on December 28, 2024 and last revised on June 3, 2025. That is the authors' own claim from a specific period and setup. This post did not rerun the experiments, and the paper's abstract does not tell you how the framework will behave on today's market or with today's models.
What Hacker News skeptics say
The repo was discussed on Hacker News with about 120 points and 81 comments. The thread is a good checklist of the concerns, presented here as commenters' reports and opinions, not verified facts:
- Does specialization help? hacker_9, who says they worked in hedge funds for a decade, argues that LLM agents share one brain, so splitting them into specialists adds little, and that browsing Reddit for sentiment is not a real trading input since such signals lag.
- Code issues. dsl said they spent about an hour in the code and listed concerns, including Yahoo News being fed to two analysts and double weighted, a sentiment prompt that primes bullishness on one stock, and a parsing bug in the memory loop. These were reported against the version at the time, and later releases list fixes in similar areas, but this post did not verify each one.
- Where are the results? jatins looks for a performance section or at least a backtest, and ctdonnet asks where the test results are that show the tool fulfills its purpose.
- A live experiment. pdp said their own LLM trading experiment had lost a little money so far, with small losses.
- Costs. Several commenters joked about token costs versus profits. It is a fair practical point: a full debate with many rounds multiplies API calls.
- Single agent or many? shaolinspirit asks whether a single agent with a good harness would do better. The README does not settle that, and the repo has not published a controlled comparison in the pages reviewed.
Not every reaction was negative. petesergeant argued that multi-agent workflows with different underlying LLMs produce better results, at least for code review. That is a claim about code, not trading.
Good uses and bad uses
| Use | Verdict |
|---|---|
| Studying a multi-agent graph with real roles, memory and checkpoints | Good. The code is a readable LangGraph example |
| Prototyping research-note generation with citations to filings | Reasonable, with human review |
| Testing which models handle long debates | Reasonable |
| Automatically placing real orders | Not supported by the README and a bad idea without independent validation |
| Treating output as advice | The project states it is not financial advice |
Safety, cost and legal cautions
- Costs scale with rounds. The README exposes
max_debate_roundsandmax_risk_rounds; set them low while testing. - Credentials. Keep API keys in
.env, not in a repo. If you use a plan sign-in, protectchatgpt-auth.json. - Data terms. The README notes SEC EDGAR asks callers to identify themselves with a contact address; set your own.
- Agents that act. If you extend the framework to connect to a broker, treat it like any tool-using agent: restrict permissions and log actions. AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
For another trending agent repo with a clear safety angle, see explainx.ai's guide to PentAGI and the iFixAI agent audit.
Bottom line
TradingAgents is popular because the idea is easy to picture: a small firm of LLMs arguing about a stock. The repository is actively maintained, with frequent releases and a serious focus on look-ahead bias. What it has not offered, in the sources reviewed, is independent evidence that the debate produces an edge. Use it as a learning project for multi-agent design and for point-in-time data hygiene, and treat any profit claim as unproven.
Related reading
- Multi-agent orchestration patterns
- Multi-agent error propagation patterns
- OpenAI Decisions API and GPT-6 Luna at DevDay
- What are agent skills? A complete guide
- PentAGI: autonomous pentest agent trending on GitHub
- iFixAI agent audit
Details are accurate as of October 25, 2026. Star counts and release notes change quickly. This is not financial advice.