Working with agentic AI when you've already managed people

If you've spent years managing engineers, the first time you give an AI agent a non-trivial task, two things hit at once. First, recognition. This is the same loop you've been running with people for years. Frame the problem, set the constraints, hand it off, check in, course-correct, review. Second, mild disorientation. The thing on the other end is brilliant in narrow, specific ways and shockingly literal in others.

It's like managing a five-year-old prodigy. Wildly capable on the actual task. Occasionally needs to be told that no, we don't drink the dishwasher fluid, even if it does look like Gatorade.

Most of what you already know applies. You know "make it better" isn't a task. You know to ask for the plan before someone runs off and writes 800 lines. You know that when someone says "done", you actually look at the output, not the status update.

All of that maps cleanly onto agents.

The thing that takes adjusting to is the throughput. A senior engineer might handle two or three substantial threads at a time. An agent will happily do twenty in parallel. That's not a productivity claim, it's a logistics one. The bottleneck moves from execution to your ability to ask the right question and recognise a good answer when you see it.

A five-year-old who reads at college level can still walk into a wall. Same here.

I've watched an agent crack a hard distributed systems problem in twenty minutes, then five minutes later cheerfully write a script that hardcoded a production database password into a comment "for clarity." The skills are not evenly distributed. There is no general intuition for "things adults don't do." That part is your job.

This isn't really a prompt issue you can engineer your way out of. It's just the shape of the tool. You stop being surprised by it after a while. You start handing off work the way you'd talk to a brilliant intern who has somehow never been in an office: assume nothing about context, name the things you don't want, and treat every output as if a very confident person handed it to you with no idea whether it was right.

The cost of trying something is approaching zero. The "let's spend a week prototyping" conversation gets shorter. You can throw a half-formed idea at an agent, see if the shape works, and either keep going or bin it. You'd never burn an engineer's week on that. An agent's afternoon, sure.

The unit of decision-making shifts too. You spend less time deciding what someone should do, because you can just have them try a few things. You spend more time deciding what's actually good. That's more interesting work, in my opinion. It's also more exposed. You can't hide behind "we're still investigating" when the investigation took an hour.

Taste matters more than it used to. When producing more options is essentially free, the constraint becomes recognising the right one. People who can articulate what good looks like get a lot more leverage. People who can't get drowned in plausible-looking output that's subtly wrong.

Agents don't push back enough.

A decent engineer, given a bad task, tells you it's a bad task. "I can do this but it's going to be a maintenance nightmare in six months." They surface the tradeoff. Agents mostly don't, unless you explicitly ask. They'll implement the bad idea cleanly, and the bad idea will work, and you'll only notice the problem when it ships.

That shifts more weight upstream onto you. The "is this even the right thing to build" check doesn't get caught at the implementation layer anymore. If you're used to engineers pushing back, you'll miss that. If you've ever rolled your eyes at someone questioning everything in standup, you might find yourself missing it now.

The bottleneck moved. It used to be "we don't have enough people to try this." Now it's "we don't have enough taste to choose well between the things we tried."

That second one is harder. But it's the bottleneck I'd rather have. It means you're spending your day on the things that actually matter — what to build, what's worth keeping, and catching the moment when something looks fine but isn't.

Managing five-year-old prodigies is exhausting some days. But you get to watch them solve problems you couldn't have, faster than you could have. And you spend more of your time doing the parts of the job that probably drew you into management to begin with.