Full autonomy is rare, and that's usually correct
It's tempting to think of a 'finished' production agent as one that runs completely on its own with zero human involvement. In practice, fully autonomous agents making high-stakes decisions with no human checkpoints at all are rare, and not because the technology can't do it, but because the cost of an unnoticed bad decision is often much higher than the cost of a small delay for review. A support agent that occasionally sends an oddly worded reply is low stakes. An agent that approves a refund, modifies a production database, or sends an email on a company's behalf is a different risk category entirely, even if it's right the vast majority of the time.
The realistic model most production systems converge on is graduated autonomy: the agent operates independently on tasks that are low-risk and well understood, where the failure mode is minor and reversible, and it routes anything unusual, high-value, or low-confidence to a human for review before acting. This isn't a compromise or an admission that the agent isn't good enough, it's a deliberate design choice that matches the level of oversight to the actual stakes of each decision.