Agents run tools in a loop.

The ability to test their work reliably is a tool, if you don't give them that, it's kinda silly to expect any kind of quality output.