Evaluating AI agents requires tracking multi-step tool calls in live environments

1. Define a sequence of tool calls that represent a complete user task.
2. Deploy the AI agent in a sandboxed live environment to test real-world execution.
3. Track the success and failure rates of each sequential step in the workflow.
4. Analyze the specific tool calls where the agent deviates or fails to complete the task.