I don't get what this is about.

The examples they show are someone typing something into a prompt and then some agent starts coding.

But the hard part in automation these days is not writing software. It is "how does the software get the credentials to access data and services? What APIs does it use?". So before being able to type a prompt and then use the software created, there must be the hard part: Setting up all those interconnectivity pieces. Making sure the software can access and change data, trigger actions, but only within a certain scope.