Autonomy you can check.
Agent demos look effortless. Daily use is where they break. Reliable autonomy comes from plain disciplines: plan the task, set the limits, check the result.
An agent that books a flight in a demo video and an agent you would let book your flight are different things. The first needs to succeed once. The second needs to succeed on a slow Tuesday, on a site that changed its layout last week, with your card and your name on the line.
When autonomous agents fail in practice, they tend to fail in a few recognisable ways. They stop halfway and report success. They act confidently on a wrong assumption. They loop on a page they do not understand. Or they do something consequential that nobody approved.
Plan first, then act
Before an agent touches a website, it should know what finished looks like and what the steps are likely to be. A plan makes failures visible: if step three was "choose a slot" and the agent is now on a payment page, something went wrong.
Small operational rules matter here more than clever models. One we learned the hard way: on most shopping and booking sites, sign in first, then add to the cart. Many sites let you build a cart as a guest and then ask you to log in at checkout, which can empty the cart or lose the saved address. Signing in first avoids a whole class of failures.
Draw the boundary before you need it
Some actions are cheap to undo and some are not. Reading a page, filling a form and searching for options can run on their own. Paying, booking, cancelling and sending a message to another person should wait for a yes, with the exact details shown: the amount, the date, the recipient.
Autonomy is not the absence of approvals. It is knowing exactly which steps need one.
Credentials deserve their own boundary. Passwords and card details belong in a vault that the agent can use without the model ever seeing them. When a site asks for a one-time password, the agent should ask you for it in the chat, use it, and move on.
Check the result
The final click is not the end of a task. The end is evidence that the task worked: a confirmation number on the page, a confirmation email in the inbox, an event on the calendar. An agent should look for that evidence before it tells you anything is done.
When it cannot find the evidence, the honest report is "I could not confirm this", along with what it saw. That is less satisfying than "Done!", and far more useful.
Know when to hand over
Some tasks will not complete. A site blocks automated access, a form needs a document you have not shared, an option requires a judgement only you can make. A reliable agent recognises this early and hands over cleanly, with a direct link to the page and a summary of where it stopped, instead of trying the same thing ten times.
None of this is glamorous. It is the difference between an agent that impresses you once and one you hand the next task to without thinking about it.