Every team is shipping agents. Almost nobody can tell you if they work. Page views and funnels were built for humans clicking buttons, and they break when the user is a model making hundreds of decisions per task. This talk covers the signals that actually predict agent success (task completion, retries, human overrides, cost per outcome) and how to capture and query them at scale without drowning in traces.
Rafael Aviles is Co-Founder and CEO of ObsessionDB (YC F26), a real time data warehouse for AI agents. He previously ran a profitable data infrastructure company at petabyte scale that handled billions of real time API requests every month. Earlier he consulted on analytics for Wendy’s, Groupon, FOX, and ESPN and other large global enterprises.