video1.0<iframe src="https://www.loom.com/embed/1bdce19966af424fa1fb17cfbe73e5c8" frameborder="0" width="1662" height="1246" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>12461662Loomhttps://www.loom.com12461662https://cdn.loom.com/sessions/thumbnails/1bdce19966af424fa1fb17cfbe73e5c8-925be118b8429e42.gif263.445333This Loom explains a recursive verifier engine and how it decides whether an agent’s work is truly done. The system gives an agent a one-line goal and a token budget, then builds a tree of plan, checks, falsify, work, verify, and rerun steps to produce a strict “done” or “not done” verdict. It prevents false success by rerunning every check with a model family that contributed nothing, and it reports exactly which check failed. The creator highlights a real failure case where deleting data older than 30 days was reported done despite only testing code the agent wrote, not the unknown external system configuration.