Open-source TradingAgents framework assigns market research to teams of LLM agents

# Open-source TradingAgents framework assigns market research to teams of LLM agents

*Event date: 2026-09-08*

Tauric Research has released TradingAgents as an open-source framework for experimenting with financial analysis performed by teams of large-language-model agents. The project imitates the division of work inside a trading firm, assigning separate roles to fundamental, sentiment and technical analysts before passing their findings through debate, trading and risk-management stages.

The design is intended to let specialized agents examine the same market from different perspectives and discuss a proposed strategy before producing a decision. TradingAgents is built with LangGraph, which organizes the roles and the flow of state between them. Users can select a ticker, analysis date, language-model provider and research depth through the command-line interface or initialize the framework as a Python module.

The software supports hosted model providers as well as local and enterprise configurations. Its documentation lists OpenAI, Google, Anthropic, xAI, DeepSeek, Qwen, GLM, MiniMax, OpenRouter, Azure OpenAI and AWS Bedrock among the available back ends. It can also connect to local models through Ollama or to OpenAI-compatible servers such as vLLM, LM Studio and llama.cpp. Some configurations require accounts and credentials from those providers, while a local endpoint can operate without a provider key.

TradingAgents can work with markets covered by Yahoo Finance, including exchange-qualified ticker symbols. According to the project, it resolves a company's identity from its ticker and grounds exact price and indicator statements in a verified data snapshot. Those controls address two recurring risks in automated market reports: analyzing the wrong company and generating price levels not found in the underlying data.

The framework keeps a decision log across runs. After a completed analysis, it appends the decision to a Markdown memory file. On a later run for the same ticker, it can retrieve the realized return, compare it with an alpha benchmark and generate a short reflection. Recent decisions and cross-ticker lessons are then supplied to the portfolio-manager agent. An optional checkpoint feature stores LangGraph state in per-ticker SQLite databases so interrupted analyses can resume.

The developers emphasize that results are not deterministic. Model sampling can vary even when settings appear unchanged, and live news or social sources can change between runs. Fixing an analysis date can stabilize the price and indicator window but does not freeze every external input. Backtest outcomes may therefore differ with the selected model, temperature, time period and data quality.

That warning is important because the project does not present itself as an investment product or a strategy with guaranteed returns. Its maintainers describe it as a research scaffold and say its output is not financial, investment or trading advice. The multiple-agent structure may make competing assumptions visible, but debate between language models does not independently validate their conclusions. Anyone evaluating the framework still needs to inspect its data sources, prompts, model behavior and risk controls rather than treating the final decision as a reliable trade signal.