Evaluating AI agents requires measuring multi-step reasoning in live environments

1. Define a set of multi-step tasks that require sequential tool calls in a sandboxed environment.
2. Monitor the agent's tool selection accuracy and error recovery at each step of the execution chain.
3. Measure the overall task completion rate and compare it against baseline performance metrics.