A separate judge checks whether a long goal was met
2026-09-05
Giving an agent a big goal was always easy. The hard part was the agent still holding that goal a hundred conversations and a dozen restarts later, and not reporting success where there was none. Every long goal now carries its own control. The agent turns the goal into a measurable condition and registers it, so it stays in front of the agent on every turn and survives context compaction, restarts and changes of subject. Acceptance is done by a separate judge that did not do the work and looks at it with a clean context, from evidence only: it runs the checks in the agent's sandbox, looks into the CRM, the spreadsheets, the pages and the chats, and reads the working log. The agent's own claim of done is not proof. If the agent stalls, the judge pushes it with a concrete next step; if it is genuinely waiting on the client for an answer, a file or an access, that wait is recorded instead of the person being nagged again. A correction in one sentence changes course, and when the deadline passes the goal closes with a report rather than hanging. The result reaches you from the agent in the same chat: what was achieved and where to see it.