Agents are now fully linked to software repositories and on their own, typically on a schedule, will look at the code and suggest improvements based on an operators request. What’s missing now (and it’s a tiny one) is a sense of memory on how the agent has acted in the past, and a sense of acting only when it makes sense. At the moment many of these systems will make a change no matter how useful, and if it’s only marginally useful it might find that undoing the same change the next day is also an improvement!

It can be - however the results are really not that great unless you have some serious computing power in my experience!