i stopped judging these agents by what they do in a demo and started counting how many of my open loops they close

Most of the agent talk here measures the tool on task success rate or how many integrations it lists. neither predicted whether i'd actually keep one open the next week. the number that did: how many of my open loops close without me touching them. the action item that dies between granola notes and linear, the follow-up that gets drafted but never sent, the hubspot field nobody updates. That gap is where the week leaks, not the meeting itself. The one desktop thing that moved it for me was runner, mostly because it pulls context across gmail, calendar and the tracker in a single task and asks before it writes anything. ships like 31 workflow templates out of the box but i only ever used the follow-up one. the connector count told me nothing, one loop closing on its own told me everything. If i had to pick one stat to judge these by now, it's 'tasks i didn't have to re-key into the system of record.' every benchmark i've seen scores the demo, which is the part of the job that was never the problem. written with ai