Grading the Machine: Can You Prove What Your AI Agent Actually Did?

Every SaaS company now has an AI agent doing real work — closing tickets, touching orders, making calls a human used to make. When something goes wrong, most founders discover they can’t actually answer the first question anyone asks: what did the agent do, and why? Ding Yuan has spent his career on exactly this problem — first as the researcher U of T made a Canada Research Chair for, now as the founder building tools to make agent behavior visible and provable at massive scale. In this fireside, Mike Potter — who’s spent a decade building the literal undo button for SaaS data at Rewind — presses him on what it actually takes to trust an autonomous system with real authority: what “good enough” means when no human is checking the work, why most companies don’t find out their agent broke something until it’s too late, and what founders need to build in now, before they hand an agent something they can’t take back.