Deploy a full AI trading team — analysts, bull/bear researchers, trader, risk manager, and portfolio manager — to collaboratively evaluate any ticker, debate strategies, and produce auditable trading decisions. All open source.
tradingagents analyze --ticker NVDA --date 2026-01-15 --checkpoint
analyst lenses in one run
research sides in structured debate
persistent decision memory per ticker
markets through Yahoo Finance tickers
TradingAgents is an open-source framework that replaces a single model with a team of specialized LLM agents — each with a focused role, connected by LangGraph, and coordinated through structured debate and multi-layer risk review.
Instead of asking one model to do everything, TradingAgents assigns narrow roles to specialized agents: four analysts (fundamentals, sentiment, news, technical), two researchers (bull/bear), a trader, a risk manager, and a portfolio manager. Each agent is powered by an LLM and connected through a LangGraph state machine that tracks every decision step.
Before any trade is proposed, bull and bear researchers engage in structured multi-round debates, cross-examining evidence from the analyst team. This adversarial process reduces single-model hallucination, surfaces blind spots on both sides, and produces conclusions that are more balanced than any one model could achieve alone.
Every trade proposal passes through a dedicated risk management agent that evaluates volatility, liquidity, and portfolio exposure — then through a portfolio manager with final approval authority. No simulated execution happens without passing both gates.
TradingAgents works with any Yahoo Finance ticker: US equities (AAPL, SPY), Hong Kong (0700.HK), Japan (7203.T), London (AZN.L), India (RELIANCE.NS), China A-shares (600519.SS), and crypto pairs (BTC-USD, ETH-USD). Company identity resolves deterministically before any analysis begins.
TradingAgents mirrors a real trading firm. Instead of asking one model to do everything, it assigns narrow, auditable roles to nine specialized agents organized across three layers — analysts, researchers, and execution.
Evaluates financial statements, earnings quality, and intrinsic value signals. Flags red flags in revenue recognition, margin trends, and balance sheet health.
Aggregates news tone, StockTwits sentiment, and Reddit discussion into a short-horizon market mood indicator. Surfaces crowd psychology that fundamentals alone miss.
Monitors global macroeconomic indicators and market-moving headlines. Interprets how regulatory changes, earnings surprises, and geopolitical events affect the ticker.
Applies indicators such as MACD, RSI, and moving averages to detect patterns and forecast price direction. All claims are grounded in a verified data snapshot.
Builds the upside case from analyst findings. Pressure-tests where conviction should increase and what catalysts could drive further gains.
Builds the downside case. Surfaces hidden fragility, valuation risks, and scenarios where the bull case breaks down.
Synthesizes research and debate outcomes into a structured action proposal with recommended position, timing, and rationale.
Re-evaluates the proposal through volatility, liquidity, and portfolio exposure lenses. Can reject or request modifications before escalation.
Makes the final approve-or-reject decision. Approved proposals are sent to the simulated exchange; all decisions are logged for audit.
Three entry points into the project — choose the one that matches how you prefer to explore a new research framework.
Clone, install, and run. Includes CLI commands, environment setup, provider configuration, and release tags.
Read the research paper for system architecture, role design, and evaluation methodology before diving into code.
Stay on this page for a dense walkthrough of the architecture, workflow, and key design decisions.
Decision flow
Every stage has a narrow job and passes a structured artifact to the next stage. The result is a decision path you can audit, resume, and compare across runs.
Choose ticker, date, LLM provider, model pair, and research depth — all configurable from the CLI or a Python config object.
Four specialist analysts independently collect fundamentals, sentiment, news, and technical context using market-aware ticker resolution and verified data snapshots.
Both researchers challenge the analyst findings from opposing sides. Multiple debate rounds surface evidence gaps before any trade is proposed.
The trader synthesizes research and debate into a structured recommendation: action, position size, timing, and supporting rationale.
Risk management evaluates volatility and exposure. The portfolio manager approves or rejects. Approved proposals execute against the simulated exchange and are written to persistent memory.
Verified data-access contracts reduce fabricated market claims.
Grounded analysts and deterministic ticker resolution tighten factual drift.
Checkpointed LangGraph runs resume from the last successful step — no restarts.
The framework stays flexible where experimentation matters and disciplined where factual accuracy is critical.
Agents cite verified data snapshots rather than loosely inferred price levels. Ticker identity is resolved deterministically before any analysis begins.
Use frontier APIs, local models, or any OpenAI-compatible endpoint — no graph changes required.
Any Yahoo Finance market works with exchange-suffixed tickers. The same workflow spans equities, ADRs, and crypto.
Two state layers keep long-running research sessions usable instead of disposable.
The project is explicit about what varies between runs so researchers can separate signal from noise.
Think of TradingAgents as a constrained research pipeline: clearly defined inputs flow through observable intermediate states to structured outputs, with explicit limits on reproducibility.
A symbol with optional exchange suffix (e.g., 0700.HK, 600519.SS) plus a target trade date anchor the entire run.
Swap between frontier APIs, local Ollama models, or any OpenAI-compatible backend without touching the workflow graph.
More debate rounds yield deeper cross-examination at the cost of runtime. Checkpoint mode protects partially completed work from crashes.
Fundamental, sentiment, news, and technical findings compiled into a structured evidence base — not a single opinion.
The trader emits a specific action (buy, sell, hold, pass) with position size, timing, and rationale instead of a vague market narrative.
Completed runs write to a decision log for cross-session learning. Per-ticker checkpoints let you resume interrupted runs from the last successful node.
We evaluated TradingAgents using a comprehensive backtesting simulation from January 1 to March 29, 2024, across major technology stocks including Apple (AAPL), Nvidia (NVDA), Microsoft (MSFT), Meta (META), and Google (GOOGL). Agents made decisions based solely on data available up to each trading day, eliminating look-ahead bias. The multi-modal dataset integrated historical stock prices, news articles, social media sentiment, insider transactions, financial statements, and 60 technical indicators per asset.
Results depend on model choice, temperature, data quality, and time window. For research demonstration only — not investment advice. Past performance does not guarantee future results.
Build and run
Start from the CLI for interactive exploration, use the Python package for scripted experiments, or run in Docker for an isolated environment.
Best for interactive runs and quick provider or model switching. Launch the TUI and configure everything from one screen.
git clone https://github.com/TauricResearch/TradingAgents.git
cd TradingAgents && pip install .
tradingagentsBest for scripted experiments and integrating the multi-agent graph into a broader research pipeline.
from tradingagents.graph.trading_graph import TradingAgentsGraph
ta = TradingAgentsGraph(debug=True)
_, decision = ta.propagate("NVDA", "2026-01-15")Best for isolated environments, team sharing, and Ollama-backed local inference without dependency conflicts.
cp .env.example .env
docker compose run --rm tradingagents
docker compose --profile ollama run --rm tradingagents-ollamaTradingAgents occupies a specific space between academic research and simple bot wrappers. This comparison shows how it differs from both.
| Dimension | TradingAgents | Trading-R1 paper | Single-agent bot |
|---|---|---|---|
| What it is | Runnable open-source framework with CLI, Python API, Docker, and a LangGraph workflow graph. | Research paper describing system architecture, role design, and evaluation methodology. | Typically a prompt wrapper around one LLM with no structured process. |
| Decision process | Multi-stage pipeline: four analysts → bull/bear debate → trader → risk review → portfolio approval. | Explains and motivates the multi-agent reasoning approach but is not itself executable. | Single model produces a recommendation with no internal debate or adversarial review. |
| Auditability | LangGraph state tracking, checkpoint resume, persistent decision log, and narrow role outputs. | Useful for conceptual understanding of audit principles but no executable audit trail. | Hard to inspect because all reasoning is compressed into one opaque model response. |
| Flexibility | 10+ LLM providers, local models, global tickers, Docker, and configurable debate depth. | Not an operational tool; supports framework interpretation and research context. | Often locked to one model, one market, or one API with minimal configuration options. |
Recent versions focused on making the framework more grounded, more resumable, and easier to run across providers and regions.
Verified data-access contract, expanded provider registry, FRED and Polymarket vendors, and CI gate.
Grounded Sentiment Analyst, GPT-5.5 coverage, dual-region Qwen and GLM support, remote Ollama, and benchmark expansion.
Structured-output agents, LangGraph checkpoint resume, persistent decision log, Docker, and Windows UTF-8 fix.
Multi-language support, unified model catalog, OpenAI Responses API, and newer model families across providers.
Full multi-provider LLM support across GPT, Gemini, Claude, and Grok families.
The research report landed first, with Terminal expected to follow.
Practical answers to recurring questions about who maintains TradingAgents, what the outputs mean, and how to evaluate the framework.
TradingAgents is maintained by Tauric Research and released under Apache 2.0 — fully open source, commercially usable, and auditable. The GitHub repository is the authoritative source for code, issues, and documentation.
Evaluate the framework on structural rigor: grounded data access, specialist role separation, adversarial debate before execution, checkpoint recovery, and the explicit distinction between research output and financial advice.
A trading bot typically places automated orders based on signals. TradingAgents is a research framework that produces structured analyses and simulated decisions — it does not connect to brokerages or execute live trades out of the box.
As structured research on a specific ticker and date, not as trading signals. The framework helps you inspect a decision path and compare reasoning across runs — it does not remove market uncertainty or model variance.
Straight answers about what this framework can and cannot do.
Install in minutes, run your first ticker analysis, inspect the decision path, then swap providers, models, and markets as you go.