TradingAgents: Can AI Agents Really Run a Trading Firm?
By Matt Cullerton, founder of Mavric and Apso
TradingAgents is an open-source framework from Tauric Research that runs a team of LLM agents (analysts, bull and bear researchers, a trader, risk managers and a portfolio manager) to produce a buy, sell or hold recommendation for a stock. A TikTok from @ray_fu presents it as something "someone just built" that works like "an entire Wall Street trading firm." The agent design is described accurately, but the project dates to December 2024, it does not execute trades, and its only published returns come from a three-month backtest.
What is TradingAgents?
TauricResearch/TradingAgents is an Apache-2.0 project based on the paper TradingAgents: Multi-Agents LLM Financial Trading Framework by Yijia Xiao, Edward Sun, Di Luo and Wei Wang. The paper went up on arXiv on 28 December 2024, the same day the repository was created. The video was posted on 31 August 2026, about 20 months later.
It is widely known. On 5 October 2026 the GitHub API reported 109,820 stars and 21,114 forks.
How do the TradingAgents agents work?
The video's description of the pipeline matches the paper and README:
- Four analysts work in parallel. A fundamentals analyst evaluates company financials, a sentiment analyst aggregates "news headlines, StockTwits, and Reddit chatter," a news analyst tracks global news and macro indicators, and a technical analyst uses indicators "like MACD and RSI."
- Bull and bear researchers debate the analysts' findings.
- A trader agent proposes a plan.
- A risk management team can reject or revise it.
- A portfolio manager approves or rejects the final proposal.
The repo also has a memory module, and the paper describes agents reflecting on past decisions against real market outcomes.
Does TradingAgents place real trades?
No. The video shows a filled order ticket with share quantity, order type, stop, target and position size. The README describes the output as a buy, sell or hold recommendation with a confidence rating, evaluated against a simulated exchange. The tradingagents/ package includes agents, dataflows, graph, llm_clients, memory, backtest.py and portfolio.py, but no broker adapter or order-routing code. Connecting it to a real broker is work you would do yourself. Third-party forks such as AlpacaTradingAgent exist for that reason.
The data sources are free and freemium feeds: Yahoo Finance and Alpha Vantage for prices, SEC EDGAR for fundamentals, FRED for macro data, and StockTwits, Reddit and news aggregation for sentiment. The "Options Flow" desk shown in the video has no counterpart in the project.
Is the app in the video real?
No. It is a designed mockup styled like Claude, with a live ticker tape, a "6 desks analyzing AAPL" panel and a trade ticket. Claude has none of those. Anthropic's Claude for Financial Services connects data from LSEG, FactSet, S&P Global, Morningstar and PitchBook for work such as portfolio and compliance reviews. It does not place orders.
The central example, "Orbit Dynamics (ORBT)," is invented. ORBT is the real over-the-counter ticker for Orbit International Corp., an electronics manufacturer in Hauppauge, New York, founded in 1957.
How well does TradingAgents perform?
The video quotes no returns. The paper reports a backtest from 1 January to 29 March 2024 against five baselines (Buy and Hold, MACD, KDJ+RSI, ZMR and SMA):
- AAPL: 26.62% cumulative return, Sharpe ratio 8.21, 0.91% maximum drawdown
- GOOGL: 24.36% cumulative return, Sharpe ratio 6.39, 1.69% maximum drawdown
- AMZN: 23.21% cumulative return, Sharpe ratio 5.60, 2.11% maximum drawdown
The authors say the test ran only three months "due to intensive LLM and tool use" and note there were few pullbacks in that period. A Sharpe ratio above 6 over about 60 trading days on three large tech stocks during a rally shows the pipeline works. It does not show a lasting edge. No audited track record exists. One unreviewed replication repository, created on 5 October 2026, claims the AAPL result does not reproduce.
Running it also costs money. The paper used gpt-4o-mini and gpt-4o for quick roles and o1-preview for reasoning-heavy ones, and model spend is the reason the authors gave for the short test.
Do you need a license to run an AI trading firm?
Trading your own money generally does not. The SEC's broker-dealer guide says "individuals who buy and sell securities for themselves generally are considered traders and not dealers."
Other people's money or advice changes that:
- Selling advice or signals. An investment adviser provides "advice to others or issuing reports or analyses about securities" for compensation and must register with the SEC or state regulators unless exempt. Advisers under $25 million in assets register with states, and SEC registration is required at $110 million.
- Executing trades for others. Section 15(a)(1) makes it unlawful for an unregistered broker or dealer to effect securities transactions, and FINRA Rule 1210 requires individual registration. The SEC's 2023 amendments to Rule 15b9-1 narrowed a proprietary trading exemption.
- Futures, forex or swaps. The NFA requires registration as a commodity trading advisor or commodity pool operator.
Anthropic's Usage Policy also treats investment advice as high-risk and requires that "a qualified professional in that field must review the content or decision prior to dissemination or finalization." The TradingAgents README says the framework "is designed for research purposes" and "is not intended as financial, investment, or trading advice."
Bottom line
TradingAgents is a real, popular research framework whose multi-agent design the video describes well. It produces recommendations against a simulated exchange, has no audited returns beyond a three-month backtest, and turning it into a "firm" brings registration duties the video never mentions. The video's "comment firm" offer points to a public repository anyone can clone.
Sources
- TauricResearch/TradingAgents on GitHub - the project
- TradingAgents README - agents, data sources and research disclaimer
- arXiv:2412.20138 - the paper and revision history
- arXiv:2412.20138, full HTML - backtest results and limitations
- AlpacaTradingAgent - a third-party fork with broker integration
- M0-AR/trading-agents-falsification-2026 - an unreviewed replication attempt
- SEC Guide to Broker-Dealer Registration - trader vs dealer
- SEC: Investment Advisers - adviser definition and thresholds
- SEC Release 34-98202 - Rule 15b9-1 amendments
- FINRA Rule 1210 - registration requirements
- NFA: Who Has to Register - futures and forex registration
- Anthropic Usage Policy - rules for financial advice use cases
- Claude for Financial Services - Anthropic's finance product
- Orbit International Corp. (ORBT) - the real owner of the ticker in the demo
- "Every Living Breathing Moment" by Clavier - the music in the clip
Published August 31, 2026. Based on a TikTok by @ray_fu. Full video notes (markdown)