Fetching the agents and everything logged so far.
Call each number, act like a real customer, and log what happens. Every problem gets pinned to the exact node inside the agent, so a fix goes straight to the right place instead of being guessed at.
Pick the part of the call it went wrong in, then write what it actually said. Word for word beats a summary.
Every call you have logged, saved as it happens. Nobody loses work and nothing has to be typed up afterwards.
Everything everyone has logged, and how far through the plan each tester is.