Personally, if I could run capable-enough inference on hardware I control, and could rely on the harness asking me for mechanistic confirmation before the agent can take consequential actions, I'd do it immediately.