Is scale what we're discussing though?

e.g. a prompt of "fetch <article URL> and summarise it for me" is very close to what a human would be doing with a web browser, and doesn't seem to involve any kind of scaling issue.

It’s easy to write instructions that have the agent check once every fifteen minutes, or even once an hour, in perpetuity, which never sleeps. And people do write such instructions. A human can’t do that by hand for very long.

The problem is that it is hard to distinguish your one off (which seems perfectly fine) from the tidal wave of bad actors.