Better models make more things possible. They do not, by themselves, make long-lived work coherent.

George exists for the part between "I want this" and "this is actually done."

Prompts end. Processes die. Tools fail. State changes. Evidence conflicts. Implementations drift. Humans interrupt. Models make bad assumptions. Systems recover. Work continues.

A frontier model may be excellent at reasoning, coding, computer use, tools, memory retrieval, or long autonomous runs. The harder question is what keeps the original intent coherent while those capabilities are used across time, failures, changing state, multiple executors, uncertain evidence, and real consequences.

Without an architecture above those events, the human becomes the continuity layer: carrying state between sessions, remembering what is already running, deciding whether a retry duplicates work, noticing false completion, rebuilding context after failure, and repairing the execution path before the original job can continue.

The human should define the intent and the boundaries. The human should not have to become the recovery protocol.

I first felt that pressure at home, where I had become the operational memory and control point for a growing family technology estate. That origin forced George past “an assistant with tools.” The architectural problem turned out to be much broader than the household where I first noticed it.