Teaching a Multi-Agent Trader the Difference Between SPY and CSI 300

What it does TradingAgents-astock is a specialized fork of the TradingAgents multi-agent research framework. It orchestrates seven AI analysts—market, sentiment, news, fundamentals, policy, hot-money, and lockup monitoring—whose individual reports feed into bull-versus-bear debates and a final portfolio manager that issues buy, hold, or sell decisions under A-share constraints like T+1 settlement, price limits, and lot sizing. All market data is pulled from free mainland sources such as mootdx, Tencent Finance, East Money, and Sina, with no API keys or credit walls required.

The interesting bit The project treats A-share adaptation as architectural surgery rather than translation. It adds three mainland-specific analyst roles—policy (监管 and industrial guidance), hot-money (龙虎榜 and 主力 flows), and lockup (解禁 and 减持 overhang)—because the upstream’s generic four-analyst setup would miss the drivers that actually move Chinese retail-heavy markets. The LLM strategy is also tiered: fast “quick-think” models handle the seven analysts and debaters, while heavier “deep-think” models are reserved for the research manager and portfolio manager nodes where synthesis matters.

Key highlights

• Seven specialized analysts, three of them A-share native: policy, hot-money flow, and lockup expiration monitors.
• Fully localized data layer: mootdx TCP feeds, Tencent, East Money, Sina, Tonghuashun, Cailianshe, and Baidu—no Yahoo Finance, Alpha Vantage, or Tushare points required.
• Trading logic respects A-share mechanics: T+1, daily price limits, minimum lot sizes, ST rules, and benchmarking against the CSI 300 instead of SPY.
• Dual-LLM design: cheaper fast models for research and debate, heavier models reserved for the two manager nodes that write the final investment plan.
• Built-in rate-limiting and Keep-Alive session reuse for East Money scraping to avoid IP bans during batch multi-agent runs.

Caveats

• Each stock analysis consumes 30–50 LLM API calls, so a paid API key is mandatory; subscription-tier chat accounts will not work.
• The README carries a standard “research and technical demonstration only” disclaimer, explicitly stating the output is not investment advice.

Verdict Worth a look if you are building quant research tools for mainland markets and want a ready-made multi-agent reasoning scaffold. Give it a pass if you need live trade execution—the project explicitly disclaims investment advice and is framed as a research demonstration, not a production trading system.