Stop babysitting AI agents
I'm having to unlearn a lot of muscle memory with AI agents.
My default when Claude gives me a wall of output is still to read the whole thing. Line by line. Understand every decision. Respond to every point.
That used to feel like being diligent.
With agents, it turns into babysitting pretty fast.
The job is shifting a little. Not just "do the task." More like figuring out what context it needs, how it can tell if it worked, and when it should pull me back in.
If it gives me five things and one seems off, I don't need to answer the whole list. I can just reply to that one thing. It'll keep the rest in context. The other four things don't disappear, and it can circle back when they matter.
If I'm getting pulled too deep into the weeds, I can ask for the higher-level version.
I'm trying to build the habit that watching it run means I should leave. Go work on something else. When it's done and needs feedback, I'll get the notification.
That last one is harder than it sounds. I still catch myself sitting there like an idiot watching text stream by.
Small example: I was trying to get an agent to rename tabs in iTerm2. It kept checking the tab name and seeing something different than what I was seeing on screen. I told it to use a screenshot, but that only worked if the tab stayed visible.
I ended up opening a separate iTerm window, putting it off to the side, and leaving it visible. That gave it a little workspace it could inspect whenever it needed to. I moved on and let it run for another 10 or 15 minutes.
It figured it out.
The old version of me would have gone back and forth with it a few more times, gotten annoyed, and probably done it myself.
The better move was setting up a check it could keep using without me sitting there as the human test runner.
That's the part I'm trying to get better at.
Not reading everything. Not hovering over the output. Setting things up so the agent can keep checking its own work, and only pulling me back in when there's actually a decision for me to make.