Autonomy is a dial, not a badge
Marketing presents autonomy as binary; operations experience it as degrees. A useful scale runs from supervised (every consequential action approved), through scoped autonomy (defined action classes pre-approved within limits — spend ceilings, rate caps, reversible writes), to full autonomy inside a boundary (the agent runs unattended within an explicit scope, with everything logged). Most production agents live in the middle band, and the interesting governance question is never 'is it autonomous' but 'which actions, at which limits, on whose evidence'.
What earns a wider setting
Autonomy should be purchased with evidence, not enthusiasm. The currency is the agent's evaluation record on representative tasks, its incident history, and the reversibility of the next action class under consideration. Widening goes one class at a time — let it send the internal summary unattended before it touches anything customer-visible — and narrows automatically when evaluations regress or the model underneath changes. This is the autonomy-as-dial discipline from the [governance guide](/guides/govern-agentic-ai), and it is what separates an autonomous agent from an unattended one.
The failure shape autonomy buys
Supervised systems fail loudly — a human declines the bad action. Autonomous systems fail quietly and compoundingly: a wrong step becomes the premise for the next, at machine speed, until something external notices. That shape dictates the controls: exit conditions the agent cannot argue with (step budgets, spend ceilings, terminal states), [runtime gates outside the agent's own reasoning](/guides/secure-agentic-ai), tracing rich enough to reconstruct any run, and a kill switch that has actually been rehearsed. None of this limits what autonomy is for; it is what makes the wider settings survivable.
Reading 'autonomous' in the wild
When a product calls itself an autonomous agent, translate before adopting: what can it write to, what stops it, and what would you see afterwards. Many autonomous agents are scoped automations with good branding — fine, and easier to govern than the label suggests. A few are genuinely open-loop systems holding broad credentials — also fine, if and only if the dial, gates, and evidence exist. The label predicts neither; the [classification questions](/compare/agentic-ai-vs-ai-agents) do.