Autopilot removes the hard stops: the input gate, IR approval, and build approval become recorded ledger decisions instead of blocking prompts. Escalations ride to the final report for post-hoc authentication rather than interrupting the run.
The honest caveat: the input gate is about intent, which downstream verification cannot recover — auto-approving an ambiguous feature proceeds on the model's best-guess reading, and the frozen suite will then verify that guess, confidently. Build-gate auto-approval is safer by construction: approval there just advances to RHA healing. Every auto-decision is ledger-recorded with its actor — an unattended run never records that a human authenticated anything.