ReAct Agent Evaluation — Reasoning & Action Accuracy

I will build an evaluation framework for your ReAct agent — measuring reasoning quality (do the thoughts logically justify the action selection?), action accuracy (does the agent select the correct tool and parameters?), task completion rate, reasoning efficiency (number of steps to completion), and error recovery quality (does the agent recover from failed tool calls through correct re-reasoning?). ReAct agent evaluation requires reasoning-specific metrics beyond simple task completion rate — an agent that completes tasks through blind tool-use trial and error rather than principled reasoning is fragile; evaluating reasoning quality identifies this pattern and enables targeted improvement of the system prompt and tool descriptions that drive reasoning quality.

The framework covers reasoning quality scoring, action accuracy evaluation, task completion rate, step efficiency analysis, error recovery assessment, and a reasoning quality report with targeted improvement recommendations.