TradingAgents: a research desk of LLM analysts, and why mine only paper-trades

Repo: TauricResearch/TradingAgents · Apache-2.0

My squeeze scanner is good at finding setups and bad at explaining them. TradingAgents is the opposite kind of tool: it takes one ticker and argues about it like a small research desk. Putting the two together was an obvious experiment.

Research only. Not investment advice — and the project says the same about itself.

How TradingAgents works

It's an open-source framework built on LangGraph that gives LLM agents the roles of a trading firm:

• Four analysts run in parallel: fundamentals, sentiment (news, StockTwits, Reddit), news and macro, and technicals like MACD and RSI.
• Bull and bear researchers debate what the analysts found.
• A trader turns the reports into a proposed trade with timing and size.
• A risk team weighs volatility and liquidity, and a portfolio manager approves or rejects.

It works with several model providers, including local models through Ollama, and runs from a CLI, a Python package or Docker.

How I use it

The scanner produces a short list each evening. A batch job runs TradingAgents over those tickers overnight and produces a morning summary: which names the "desk" liked, which it rejected, and why. The full debate log for each ticker is saved, so I can read the bear case when the bull case looks too neat.

The guardrail: paper only

There's a bridge script that can turn a TradingAgents decision into an order — against a broker's sandbox API, in dry-run mode by default. That's deliberate. LLM outputs are non-deterministic; the same ticker on a different day, or with a different temperature, can produce a different verdict. The README is explicit that results depend on models, data quality and randomness.

So the rule is simple: TradingAgents is a second opinion, not an execution engine. It can promote a setup for a closer look or flag a risk I missed. It doesn't move money.

What it's genuinely good at

• Forcing the counter-argument. The bear researcher reliably finds the thing I didn't want to read.
• Reading more than I can. News and sentiment across many tickers overnight is exactly what agents are for.
• Leaving a paper trail. Every decision comes with its reasoning, which makes it auditable.

What it isn't is a substitute for a backtest. A persuasive debate is not evidence. I wrote about that trap in the 80% win rate that wasn't.