TradingAgents: open-source multi-agent LLM framework that simulates an investment committee

大模型之路 · wechat · 2026-09-16

TradingAgents, an open-source project from TauricResearch that trended on GitHub, decomposes the investment process into specialized LLM agents: fundamental, sentiment and news analysts, bull/bear researchers debating the same evidence, plus trader, risk-management and manager agents that converge on a traceable decision.

The author contrasts it with rule-based quant systems—it excels at unstructured "soft" information like filings and social sentiment—and with generic frameworks like LangGraph/AutoGen, whose value is baking the finance-domain workflow in by default.

Practical caveats: many LLM calls mean high latency and cost, data feeds must be wired yourself, hallucinated numbers are costly in finance, and live-trading compliance is nontrivial. Inherent multi-agent issues—coordination overhead, information loss, debate deadlocks—also apply. Best positioned as a research copilot, not an autopilot.

Original post →

More from coding & agent

coding & agent channel →