Agentic Browsing Breaks in Boring Ways

We have not tested this. Written from published specs, prices and documentation.

Agents that drive a browser for you demo beautifully. Find me a flight, fill in this form, pull these numbers off the supplier site. Watching one work the first time is genuinely impressive. Then you use it for a week and the failures start, and they are not dramatic failures. They are dull ones, and the dullness is the point.

What breaks, in rough order of how often you will meet it.

A cookie banner. The agent cannot see the page it needs because something is covering it, and the thing covering it is different on every site.

A login that wants a code from your phone. The agent stops, and whether it tells you clearly that it stopped is a feature of the product rather than a given.

A page that quietly changed. A button moved, a label got reworded, and a flow that worked last week now does something adjacent to what you wanted.

Something that looks finished but is not. The hardest failure to catch, because the agent reports success. It filled in the form and did not notice the validation error under the fold.

Infinite patience applied to the wrong thing. Twelve minutes spent finding a page that does not exist, because nothing told it when to give up.

None of that means the category is useless. It means the useful jobs have a shape. Give an agent work where the cost of a mistake is low, where you can see the result at a glance, and where the site is one it will see again and again so its failures become predictable.

Three rules that make it survivable:

Never let it do the last irreversible step. Let it fill the cart, you press buy. Let it draft the message, you send it.

Ask it to report what it saw, not just what it did. Screenshots or a list of steps. Without that you cannot tell a real success from a reported one.

Expect to supervise it for the first ten runs of any new task, and treat run eleven as the first one that counts.

The gap between the demo and the week is the whole story with agents right now, and it is closing from both ends: the agents get better, and the people using them learn which jobs to hand over.

T

Tomasz Mieczkowski

Writes the ai beat for Drafted Tech. Every piece says at the top how we got the product, or that we never had it.

Comments (0)

wave

Leave a comment

wave

Comments are read before they appear. Your email address is never published.

Read next

wave

What are you looking for?

Press ESC to close.