The fourth and final rung of Boris Cherny's 4-Step AI Maturity Framework, following Step 3 – Supervised Autonomy (AI Agency CEO Model). The defining shift is who initiates work: at step 3 the human still kicks off most agent runs and checks in on progress; at step 4 the loop closes — most agents are started by Claude itself, not by a human prompt — and the operator's role collapses to two verbs: steer by intent, monitor by exception.
"you steer by intent and monitor by exception"
At scale this means hundreds to thousands of agents running concurrently (see Sub-Agents / Agent Swarm Fan-Out), with the human only pulled in when something is flagged wrong via the AI Agent Communication/Reporting Loop rather than checked on a schedule. The stated payoff is that quarter-long efforts (e.g., a migration) become something you simply kick off once and periodically check, rather than something you actively run.
Caveat: unlike steps 1–3, this stage is not demonstrated within the source video — it is asserted through anecdote and paraphrase of Boris Cherny's own writing, so treat the "hundreds to thousands of agents" claim as reported, not shown.