AI Agent for Scoring Trading Recommendations

The article describes an AI agent that scores trading recommendations by automatically executing trades based on signals extracted from authors' posts. Unlike relying on authors' self-reported statistics, the agent parses free-form text and images via MCP, deriving trading intentions. It runs on a test account with simulated executions. To optimize token usage, summary generation is delegated to a subagent, preventing heavy images from entering the global context. Metrics show that the Sonnet 5 model can poll every 15 minutes without exceeding weekly limits on Claude Max, while other models have shorter durations. The system avoids AI-biased backtesting by letting an external engine calculate returns, with the agent only making binary buy/sell decisions. The final rating is based on a comprehensive table of all recommended tickers, including drawdowns, Sharpe ratio, and Calmar ratio. The architecture supports local hosting via Ollama and an Anthropic-compatible proxy, ensuring weak models don't break risk management as arithmetic is handled by scripts. Trailing take and peak staleness conditions are implemented in code, and the dashboard provides charts and notifications.