Grading agent outputs
Practice describing how you built an evaluation or grading loop for an AI agent, including how you handled cases where the output looked right but was wrong.
The role wires grading into agent execution and closes the gap between agent demos and reliable production behavior.
Source posting