Over the holidays I built a little agent that does one thing: it reads the morning pile of "your cron job failed" emails, groups the ones that are the same underlying problem, and tells me which is new since yesterday. That's it. It has three tools, a model behind it, and a hard cap on how many steps it can take.
The temptation with these things is to make them do everything. Let it restart the job! Let it open a ticket! Let it page someone! I resisted all of that, on purpose. The moment an agent can take an action with consequences, you've signed up to review every action it might take, and you've lost the thing that made it worth building, which was getting five minutes of my morning back.
So it reads, it summarises, it deduplicates, and it shuts up. The "does things" in the title is doing some quiet work: the useful version of an agent doesn't grandly fix your infrastructure. It just removes the dull triage you were doing by hand, and then it gets out of the way. Small, dull, and reliable beats clever every single time.