TradingAgents Review in 2026
8 min read

TradingAgents is a free, Apache 2.0 multi-agent LLM framework from Tauric Research that analyzes one stock on one date and returns a rating with reasoning. It is a research scaffold, not a trading bot: it places no orders, runs vary, and its published returns come from a three month 2024 backtest the authors themselves flag as unusually strong.
What this covers
- TradingAgents had 121,078 GitHub stars on 7 November 2026 and is licensed Apache 2.0.
- Four analysts, a bull and bear debate, a trader, three risk analysts and a portfolio manager produce one of five ratings.
- An independent code read found no broker integration and at least 15 model calls per decision in v0.7.0.
- The paper's Sharpe ratios of 5.6 to 8.2 come from a three month Q1 2024 window the authors call unusually smooth.
- Default price data comes via yfinance, which is intended for personal and research use, so commercial use needs a licensed data source.
- Token use per run is not published; Alpha Vantage is free at 25 requests per day and paid from $49.99 per month.
TradingAgents is a free, open source Python framework from Tauric Research that runs a team of LLM agents, modeled on a trading firm, to analyze one stock on one date and return a rating with its reasoning. It is a research tool under the Apache 2.0 license, not a trading bot: the README says it is not financial advice, and an independent read of the v0.7.0 code found no broker integration and no code that places orders.
The repository at github.com/TauricResearch/TradingAgents had 121,078 stars and 23,153 forks when we fetched it on 7 November 2026, which makes it one of the most watched finance projects on GitHub. This review covers what it does, what the license allows, what it costs to run, and where it fits for a fund, a family office or a deal team.
What TradingAgents is
TradingAgents started as a research paper, "TradingAgents: Multi-Agents LLM Financial Trading Framework" by Yijia Xiao, Edward Sun, Di Luo and Wei Wang, posted to arXiv as 2412.20138. The repository was created on 28 December 2024. It is built on LangGraph, needs Python 3.11 or later, and also ships with a Docker setup.
The project moves fast. The README lists releases every month through 2026, with v0.7.0 and v0.7.1 in October 2026. Tauric Research, the team behind it, also publishes Trading-R1, a reasoning model for trading trained with reinforcement learning, which opened early access in November 2026. Tauric's own site describes a separate platform with an API and MCP server for market, filing, macro and news data.

AI for finance teams
Screen more deals without adding analysts
First system live in about six weeks. You own the code.
Book a discovery callHow the agents work
The framework splits one decision across roles that mirror a trading desk. Four analysts work in parallel: a fundamentals analyst, a sentiment analyst that reads news headlines, StockTwits and Reddit, a news analyst covering global news and macro indicators, and a technical analyst using indicators such as MACD and RSI. A bull researcher and a bear researcher then debate the analyst reports.
A trader agent turns the result into a proposal. Three risk analysts with aggressive, conservative and neutral stances argue over that proposal, and a portfolio manager makes the final call. Per the independent code review on devpik.com, the output is one of five ratings: Buy, Overweight, Hold, Underweight or Sell.
Two model tiers control cost. A quick model serves the analysts, researchers, debaters and trader, and a deep model serves the research manager and portfolio manager. Each tier can sit on its own provider, so the managers could run on Claude while the rest run on a cheaper model. The README lists GPT-6 Luna and GPT-6 Sol as the current OpenAI defaults.
You can run it from an interactive CLI, from flags for scheduled jobs, or from Python by calling TradingAgentsGraph().propagate(ticker, date). It covers any market Yahoo Finance covers, including US, Hong Kong, Tokyo, London, India, Canada, Australia, China A shares and crypto pairs. You can also pass in your current holdings so the trader and portfolio manager reason against your actual book.
What recent releases fixed
The most useful work in 2026 is about what the agents are allowed to know on a given date. US fundamentals now come from SEC EDGAR as filed. The README's example: Apple's 2008 total assets were filed as $39.6B and restated to $36.2B in 2010, so a run dated in between reads $39.6B. Yahoo statements and insider trades, which cannot be dated by publication, are withheld from past dated runs.
Each run also writes its decision to a memory log. Later runs settle past decisions against realized returns and alpha versus a regional benchmark, and the portfolio manager reads those lessons. A run_backtest function sweeps a grid of tickers and dates, enters each decision at the first close after its date, and keeps a separate memory log so live runs are not contaminated.
The published results, and how much weight they bear
The paper reports a backtest from 1 January to 29 March 2024. On AAPL it shows a 26.62% cumulative return and a Sharpe ratio of 8.21, against rule based baselines and buy and hold. GOOGL and AMZN show Sharpe ratios of 6.39 and 5.60.
Treat those numbers with care, and the authors say much of this themselves. The paper's footnote states the highest Sharpe ratio exceeds their expected range and attributes it to few pullbacks in the window. The test ran three months because each prediction used about 11 LLM calls and 20 or more tool calls.
Independent reviewers go further. Pith's review notes that the setup names five stocks but the results table shows three, with no repeated runs or confidence intervals, and that baselines are rule based rather than other LLM traders. A paper review by Kyubum Hwang adds that no ablation separates the effect of the debate structure from the effect of more model calls, and that transaction costs and slippage are not modeled.
The README is direct on this point too. It says two runs of the same ticker and date can differ, that reasoning models largely ignore temperature, that news and social sources reflect the present even in a past dated run, and that backtest results are not guaranteed to match any published figure. It calls the framework "a research scaffold," and that is the honest description.

Build library
See what we have already shipped
Working systems, the stack behind them, and what they cost to run.
Open the libraryLicense: what Apache 2.0 allows commercially
TradingAgents ships under the Apache License 2.0. The license grants a copyright license and a patent license to use, modify and distribute the code, including in commercial work. In return you must retain copyright and attribution notices, carry forward any NOTICE file, and state changes in modified files. The license does not grant use of the project's trademarks.
The code license is only half the question. TradingAgents pulls prices and statements through yfinance, whose own README says it is not affiliated with Yahoo, is meant for research and educational use, and that the Yahoo Finance API is intended for personal use only. A fund putting this into a commercial workflow needs a data source it is licensed to use that way, which usually means swapping the data layer.
What it costs to run
The software is free. The running cost is model calls plus data. Neither the README nor the paper publishes token use per run, so any dollar figure per decision would be a guess. What is known is the call count: the devpik.com review counted at least 15 model calls per decision in v0.7.0 with default settings, 13 of them on the quick tier.
| Cost line | What the sources say |
|---|---|
| Software | Free, Apache 2.0 |
| LLM calls | At least 15 per decision (devpik.com count, v0.7.0); token use per run not published |
| Model provider | Your own API key across many providers, local models via Ollama, or a ChatGPT Plus or Pro sign in that counts against plan limits |
| Alpha Vantage | Free tier capped at 25 requests per day; premium plans start at $49.99 per month |
| SEC EDGAR | No account or key needed |
| FRED macro data | Free key, optional |
| Infrastructure | Runs on a laptop or in Docker; no hosted service required |
The cost grows with scope. One ticker on one date is cheap to try. A nightly sweep across a large universe, or a backtest grid with weekly dates over a year, multiplies every one of those calls. The ReviewRamble write up makes the same point: a ChatGPT plan suits experiments, not a nightly run across hundreds of tickers.
Who should use it
TradingAgents is a good fit for researchers and students working on multi-agent systems, developers who want a well maintained LangGraph project to learn from, and quant teams who want to test whether structured debate changes a model's calls. Its point in time data handling and backtest harness are worth studying even if you never run its ratings.
It is a useful prototype for an investment team exploring how agents could draft a first pass on a name. The role structure, the bull and bear debate, and the memory log that scores past calls are all patterns that carry over into real research workflows.

Implementation services
Put AI into the workflow you actually run
Diligence, close, reconciliation, reporting — in your environment.
See how it worksWho should not
It is the wrong tool for anyone who wants software to trade their money. It places no orders, the README disclaims advice, two runs can disagree, and the published returns come from one short window. The devpik.com review also recommends keeping the sentiment analyst out of anything you backtest, because social and news sources reflect the present day.
It is also a poor fit as is for a regulated firm. Out of the box it reads public web data, sends your prompts to whichever model provider you configure, and has no access to your internal research, your models, your CRM or your deal files. Those are the things that make an investment decision yours.
Where it fits for a fund, family office or deal team
For a public markets fund, TradingAgents is best read as a reference design. The ideas that matter, separate roles, structured handoffs, a debate step, and point in time data discipline, are the same ideas you would want in a production research assistant. The gap is everything around them: licensed data, audit trails, permissions, and evaluation against your own historical decisions.
For a family office, the more common need is not a trading signal. It is consolidating custodian data, monitoring managers and drafting the quarterly view, which is a different system. See how we approach that in AI for family offices.
For private equity and deal teams, the multi-agent pattern maps well onto diligence: one agent per workstream, a skeptic that argues the other side, and a reviewer that ties every number back to the data room. That is closer to what we build in AI financial due diligence and AI deal desk systems than anything TradingAgents ships.
When an in environment build is the better answer
If you want to learn how these systems reason, clone TradingAgents and run it. It costs an afternoon and some API credit, and it will teach you more than most vendor demos.
If you want the pattern working on your own data, inside your own environment, with licensed sources and an audit trail, a custom build is usually the better answer. That is what consultance.ai does: we build custom finance AI, the first system is live in about six weeks, and you own the code. If that is the question on your desk, book a call, or browse the free prompt packs in our library first.
Sources
- TauricResearch/TradingAgents on GitHub, README, stars and forks, fetched 7 November 2026
- TradingAgents LICENSE (Apache 2.0), fetched 7 November 2026
- TradingAgents pyproject.toml, fetched 7 November 2026
- TradingAgents paper on arXiv (2412.20138), fetched 7 November 2026
- Tauric Research, fetched 7 November 2026
- devpik.com: TradingAgents read line by line, fetched 7 November 2026
- Pith review of the TradingAgents paper, fetched 7 November 2026
- Kyubum Hwang, paper review, fetched 7 November 2026
- ReviewRamble: TradingAgents GitHub repo review, fetched 7 November 2026
- yfinance on GitHub, fetched 7 November 2026
- Alpha Vantage premium plans, fetched 7 November 2026
