Mousepower: agents that can’t be measured, can’t be managed. — Maximillian Piras, Yutori

According to the description of this talk by Yutori's Maximillian Piras, a task that is hard to perform yet easy to check is a promising candidate for an agent. His framework separates uncertainty about the steps from uncertainty about what counts as success. When the steps are already fixed, a script may suffice. When success itself is ambiguous, verifying the agent's output can cost as much as doing the task. The proposed sweet spot leaves the agent room to work out a solution while giving the reviewer clear acceptance criteria. The talk also argues that measuring accepted outcomes, such as resolved support requests, says more about value than measuring generated text.