Agents, maybe, but mostly as expensive interns
I keep trying to find the shape of AI agents in my actual workflow, and I still don’t really see it.
I’ve been using opencode for a couple of weeks, and the feeling I keep coming back to is weirdly simple: this mostly feels like a very good way to burn tokens. The promise is delegation. You give the agent a job, step back, and let it carry the thing through. Obviously that sounds great.
But when I actually use them, I end up steering constantly. Clarifying. Correcting. Nudging them back onto the path I thought I’d already described. And after a while I start wondering whether I’m delegating anything at all, or just managing badly.
Maybe this is a target audience problem. If you already know how to do the task, and you already have a workflow that works, the agent can feel like extra ceremony wrapped around something you could have just done. I don’t know. Maybe this lands better for people who need a first pass, or a confidence boost, or just a way to get unstuck without holding the whole problem in their head.
What keeps nagging at me is that I’ve had more obviously useful AI moments in much smaller forms. A Vim plugin hooked up to Ollama. Good autocomplete. Spellcheck that actually catches the thing I meant. That kind of stuff feels powerful because it shortens the distance between intention and action. It helps me talk to the computer faster.
So maybe that’s the bit I believe in most. AI as interface. Not replacement brain. Not employee. Just a better surface for turning half-formed thoughts into something usable.