vignesh-codes/tradingagents

Purpose and entry points

TradingAgents is a Python, LangGraph-based multi-agent framework that produces simulated trading decisions from market data and LLM reasoning. Primary interactive entry point is cli/main.py; package/programmatic use centers on tradingagents/graph/trading_graph.py and is demonstrated by main.py. Runtime defaults are defined in tradingagents/default_config.py.

Orchestration graph

TradingAgentsGraph constructs and propagates the workflow, carrying ticker, date, reports, debate artifacts, and final decision through LangGraph nodes. Graph setup, routing/conditional progression, and inter-agent propagation are separated in:

• tradingagents/graph/trading_graph.py

• tradingagents/graph/setup.py

• tradingagents/graph/conditional_logic.py

• tradingagents/graph/propagation.py

Market-data boundary

Agent-facing tools abstract financial-data retrieval and indicator computation. The dataflow layer supports Alpha Vantage and Yahoo Finance, isolating provider-specific API calls from agent reasoning:

• tradingagents/dataflows/interface.py

• tradingagents/dataflows/alpha_vantage.py

• tradingagents/dataflows/y_finance.py

• tradingagents/agents/utils/core_stock_tools.py

• tradingagents/agents/utils/technical_indicators_tools.py

Alpha Vantage requires its API key; Yahoo Finance is another market/news source. Inputs are ticker and analysis date, with tools supplying fundamentals, prices, technicals, and news/social context.

Analyst stage

Specialist LLM agents generate independent reports on fundamentals, market/technical signals, news, and social sentiment. They are grounded by data tools and return artifacts into shared graph state:

• tradingagents/agents/analysts/fundamentals_analyst.py

• tradingagents/agents/analysts/market_analyst.py

• tradingagents/agents/analysts/news_analyst.py

• tradingagents/agents/analysts/social_media_analyst.py

Research, trade, and risk decision stages

Bull and bear researchers debate the analyst evidence; the research manager synthesizes that work. The trader proposes a transaction, then risk debaters provide aggressive, conservative, and neutral views before the portfolio manager approves or rejects the outcome:

• tradingagents/agents/researchers/bull_researcher.py

• tradingagents/agents/researchers/bear_researcher.py

• tradingagents/agents/managers/research_manager.py

• tradingagents/agents/trader/trader.py

• tradingagents/agents/risk_mgmt/

• tradingagents/agents/managers/portfolio_manager.py

Structured schemas and serialization form the important boundary between free-form model output and reliable downstream graph decisions: tradingagents/agents/schemas.py, tradingagents/agents/utils/structured.py.

LLM provider boundary

A client factory and model catalog normalize calls across OpenAI, Google, Anthropic, xAI-compatible providers, DeepSeek, Qwen, GLM, OpenRouter, Ollama, and Azure. Provider credentials are environment-driven, including enterprise settings:

• tradingagents/llm_clients/factory.py

• tradingagents/llm_clients/base_client.py

• tradingagents/llm_clients/model_catalog.py

• tradingagents/llm_clients/openai_client.py

• tradingagents/llm_clients/azure_client.py

Persistence, recovery, and deployment

Completed decisions feed a persistent memory/reflection loop, allowing later same-ticker and cross-ticker context to influence the portfolio-manager prompt. Optional per-ticker SQLite checkpoints persist LangGraph state after nodes and support interrupted-run resume:

• tradingagents/agents/utils/memory.py

• tradingagents/graph/reflection.py

• tradingagents/graph/checkpointer.py

The system runs locally as Python or in Docker via Dockerfile and docker-compose.yml; Docker Compose also supports an Ollama profile for local models.