Leave the Hard Stuff to the Humans
The pitch for AI coding tools is simple: automate the “mundane” things, so it frees you up for more interesting problems. I agree with that pitch. I just didn’t expect it to mean I’d spend my whole day doing nothing but the interesting problems, back to back, with no boring ones left in between to catch my breath.
How I normally work with it
Most of the time I scope things narrowly. Add this API endpoint to this feature. Use this format, this architecture style, follow the pattern two files over. I already know what the thing is supposed to do, and AI is excellent at that — it can follow a pattern I hand it all day long.
A few days ago I decided to try something different: speedrun a chunk of my backlog and see how far I could get with as little upfront planning as possible.
The experiment
The idea was simple. Pull the next five tasks off Jira. Spin up a separate git worktree for each one. Point an agent at each and just let it go — all the way through to a pull request — without me writing careful prompts or checking first whether the task was even something AI could reasonably finish. I wanted to see how far it would get on its own.
I ran all five sessions at once, using a session manager called herdr to manage them and ping me when one needed something. Model of choice: Opus 5.
Because none of the five tasks were narrowly scoped, they were a real grab-bag — research one of them turned into, a production deploy question on another, a small feature here, a genuinely complex one there. A couple of them touched things AI plainly didn’t have access to.
Where it kept getting stuck
Every session made real progress. It would work through as much as it could figure out on its own — and it turns out that’s a lot — and then stop, not because it was wrong, but because it didn’t have enough information to keep going without me. It needed what I’d call a hard decision: something with real tradeoffs, something that touched other people, something with no obviously correct answer sitting in the codebase or the docs waiting to be found.
So I’d stop, think it through, make the call, and hand it back. And almost immediately the next session would need me for the same reason. Multiply that by five or six running at once, and it stops feeling like supervising AI and starts feeling like an all-day rotation of hard calls with no breathing room in between.
The part nobody mentions
“Automate the boring stuff” is a fine mantra and I believe in it. I don’t want to spend my time on something AI can handle competently, and increasingly that’s most things. But that’s exactly the catch: if AI is taking every task it can see end-to-end from the codebase and the docs, what’s left over is, by definition, the stuff it couldn’t see end-to-end. The ambiguous calls. The judgment calls. The ones with no clear winner.
Run five or six agents in parallel and you’re not doing a mix of easy and hard work anymore. You’re doing hard problem, hard problem, hard problem, one after another, with the easy layer stripped out from underneath you.
I moved a lot faster that day. I also finished it more drained than I have in a long time. It’s not as simple as “AI makes everyone more productive.” It makes the repetitive, boring parts of the job nearly free — genuinely great — but it does that by concentrating everything hard onto the person left holding the decisions. Do that without pacing yourself, and it’s a pretty direct route to burnout.