TradingAgents Inspiration

TradingAgents Inspiration

This server's workflow borrows ideas from TauricResearch/TradingAgents (Apache-2.0) — a multi-agent LLM framework simulating a trading firm. Thank you to its maintainers for open-sourcing the architecture.

What we took

TradingAgents idea	Our implementation
Analyst team producing separate reports	analyst_reports — fundamentals / technical / sentiment, each timestamped
Grounded claims from a verified data snapshot	data_snapshot audit block in analyze_stock; debate prompt rule: "every claim traces to the snapshot or a fresh tool call"
Bull vs bear researcher debate	bull_bear_debate prompt — client LLM steelmans both sides from grounded data
Risk management check	debate step 4: falsification conditions, position sizing, upcoming events
Decision log with reflection	decision journal: log_decision / review_decisions with mechanical outcome verdicts

What we deliberately did NOT take (and why)

• The LangGraph orchestration stack — the MCP client is already the LLM layer; a thin server avoids per-analysis LLM cost and startup weight.
• US-centric data sources — StockTwits/Reddit have thin NSE coverage, FRED is US macro, Polymarket irrelevant. India needs different sentiment sources.
• The simulated exchange — research-only artifact; our journal + backtest harness serve the evaluation role.

Attribution

Acknowledged in the repository README and in TauricResearch/TradingAgents issue #1313.