Autonomy is a spectrum, and "autonomous agent" names its far end: software trusted to complete goals, send the email, update the record, make the purchase, with no human between decision and execution. Marketing often calls these AI employees, which is apt in both directions: they work unsupervised, and like any unsupervised hire, the question is what happens on their worst day.
Where autonomy is appropriate, it is genuinely superior: internal syncs, report generation, notifications, anything reversible and inward-facing gains nothing from a human checkpoint. The failures that make headlines share the opposite profile: outward-facing, money-touching, irreversible. An autonomous support bot inventing a refund policy is not a smarter model away from being safe; it is an architecture decision away.
When a vendor describes a product as autonomous, the useful follow-ups are: which specific actions can it take unattended, can that list be narrowed, and what is the blast radius of the worst plausible action on it? If the honest answers include "email your customers" or "move money," the alternative worth pricing is the same capability behind an approval gate, where the trade is seconds of review for structural certainty.
See the term in practice
Connect your tools read only and watch the approval queue prepare real actions from your own context. Nothing sends until you approve it.