Autonomy is the approval boundary: whether an agent needs human authorization for each action. It, rather than the interface, determines whether work can continue while people look away. section 5.2.3
Greater autonomy expands scale but also removes witnesses. Unreviewed routine changes can pass unnoticed, while failures become harder to attribute when several agents act on shared assumptions. section 5.2.3 section 5.6.7
Use autonomy as an operational decision, not a default setting:
- Keep scope conservative and automate predictable workflows before novel investigation. section 5.6.7
- Build instrumentation and observability before increasing autonomy; dashboards should precede orchestration. section 5.6.7
- Define rollback across partial, multi-agent changes rather than assuming one action can undo the whole operation. section 5.6.7
- Treat approval fatigue as a control failure: routine approvals can conceal changes that deserve scrutiny. section 5.2.3
Interface does not determine autonomy
A terminal agent can stop before writing a file, while an IDE background agent can continue without approval. Cron jobs, hooks, and pipelines make unattended execution useful; they do not by themselves remove human control. section 5.2.3













