TradingAgents

Framework extension and research · 2026 · Private research project

TradingAgents is my extension of TauricResearch’s multi-agent research framework. I added overnight research, stored evidence, decision tracking, and a separate layer of trading controls.

Open the interactive book

Research agents and trading controls

Four analysts review technical signals, sentiment, news, and company fundamentals. Their reports go through a bull and bear debate, a trade proposal, and risk review before a portfolio manager records the decision.

The AI graph produces a decision packet. Broker execution sits behind separate authorization and code-based risk checks. My repository is private.

TradingAgents extends TauricResearch TradingAgents. Four analyst reports cover technical signals, sentiment, news and company fundamentals. Bull and bear researchers debate the reports, a research manager develops a plan, and a trader proposes an action. Aggressive, conservative and neutral risk reviewers debate the proposal before the portfolio manager records a decision. Stored research context and crowd-scenario records are supporting inputs, not mandatory decision stages. MiroFish reports and local scenario records provide advisory hypotheses, not trade instructions. Saved decisions can be compared with later outcomes for reflection on a future run of the same symbol. Broker execution is separate from the AI graph: it requires authorization and code-based policy and risk checks before a paper or gated live broker action.

Keeping research connected

I built an overnight workflow that runs company analysis and saves the reports behind each decision. Supporting research records connect symbols, themes, and source references so earlier findings can be inspected again.

The video is an earlier graph walkthrough after one night of research. It shows that version of the research view, not the current system architecture.

Exploring crowd scenarios

I’m also exploring crowd scenarios through MiroFish reports and a local market-mirror module. They contribute hypotheses and questions for further research. They cannot authorize a trade.

The local module creates actor profiles and scenario records. It does not yet run a crowd reacting over multiple rounds, and I make no claim that it predicts real market behavior.

Checking predictions against outcomes

The system saves decisions so I can compare them with later outcomes and revisit the reasoning. Paper trading provides another way to test those decisions.

I’m still building and checking the evaluation process. These records do not establish that the system can trade profitably.