TradingAgents

TradingAgents is a Python package (Python 3.11 or later) with a Typer and Rich command line and a TradingAgentsGraph class. propagate(ticker, date) runs one LangGraph pipeline and returns the final state and a rating. The domain is financial trading research: agents call data tools for prices, indicators, fundamentals, news, social posts and macro series, and the output is a report tree plus a decision for the simulated exchange.

The graph has fixed roles. Selected analysts run in parallel, each with its own tools. A Bull and a Bear Researcher debate for a configured number of rounds, a Research Manager decides, and a Trader drafts a proposal. Aggressive, Neutral and Conservative risk analysts then discuss it, and a Portfolio Manager approves or rejects it. Agents talk through shared graph state, not free chat.

State is LangGraph state per run, optional SQLite checkpoints per ticker, and an append-only Markdown memory log under ~/.tradingagents/memory that feeds past decisions and reflections into later runs for the same ticker. Two model tiers are configured: a quick-thinking model for routine steps and a deep-thinking model for research and portfolio decisions.

unknown for approvals during a run: the README documents no human approval or interrupt step inside the agent graph; the Portfolio Manager is an agent that approves or rejects the proposal and sends an order to a simulated exchange. Humans choose the ticker, date, analysts, provider, models and debate rounds in the interactive CLI or by flags and TRADINGAGENTS_* variables, can supply a portfolio file so agents work against a real book, and can enable checkpoint resume after a crash.

pip install .

git clone https://github.com/TauricResearch/TradingAgents.git
cd TradingAgents
python -m venv .venv && source .venv/bin/activate   # Python 3.11 or later
pip install .
export OPENAI_API_KEY=<your-key>                     # or another provider's key

from tradingagents.graph.trading_graph import TradingAgentsGraph
from tradingagents.default_config import DEFAULT_CONFIG

config = DEFAULT_CONFIG.copy()
config["llm_provider"] = "openai"
config["max_debate_rounds"] = 1
ta = TradingAgentsGraph(debug=True, config=config)
state, decision = ta.propagate("NVDA", "2026-09-01")   # ticker, analysis date
print(decision)

• Needs Python 3.11 or later; install from a clone with pip install . (Bedrock needs pip install ".[bedrock]").
• A provider API key is required (for example OPENAI_API_KEY); the default models are GPT-6 Sol and Luna. Alpha Vantage, FRED and TypeSafe keys are optional.
• v0.5.1 reorganised the package layout and moved import paths, so code written for earlier versions may need imports updated.
• SEC asks for a contact address: set SEC_EDGAR_USER_AGENT to your own.
• Runs vary between executions because of model sampling and live news and social data; reasoning models largely ignore temperature.
• Past-dated runs withhold data sources that cannot date their figures (for example Yahoo statements and insider trades).

Researching multi-agent LLM analysis of financial markets. Agents with roles such as analyst, researcher, trader, risk debater and portfolio manager produce a decision for a ticker and date, which is sent to a simulated exchange.

The project says no: the README states it is designed for research purposes and not intended as financial, investment or trading advice, and its disclaimer says the Research is for research and educational purposes only.

Yes. The README says it was built with LangGraph, and pyproject.toml lists langgraph and langchain packages as dependencies.

No support was found in the README, CHANGELOG or repository code search, so all three are recorded as unknown.