Once I really understood how to leverage AI to get more done, I started spending a lot of time writing specs — for projects I frankly didn't have time to do myself. Suddenly those were on the table, because I had newfound ability and, in effect, workers to hand the work to.

Here's the trap I fell into: you cannot manage AI agents the way you manage people. I tried, and it was almost the exact opposite of what I should have done.

Where the instinct comes from

I was an engineer for a long time, and over the years I got promoted into leadership. For the past seven years or so I've been consulting — and honestly I'm more hands-on now than I ever was as a leader.

One thing every engineering leader internalizes: when a direct report says "I'm blocked," that's a five-alarm fire. Clearing roadblocks is one of your chief responsibilities. A lot of a leader's energy goes into making sure people don't get stuck, and into getting them unstuck fast when they do.

I had an engineer come to me saying that the prototype (and the backup) broke — it was due at the end of the week. They didn't know if they should just overnight in parts and take an extra week to build it. They wanted to update the board anyway, so they were planning to take the opportunity to order the next revision.

That's the kind of situation as a leader you can't ignore. These kinds of situations are stressful, and they require your judgment. Is there some way the prototype is recoverable? Can we keep trying to work without hardware? Should we let the customer know? Should we just pivot to the next revision of the board? And so on. You have to make a call, and you have to make it fast.

As a leader you get used to dropping everything to make those calls and balancing all the interested parties.

So when I started pushing AI agents to get more done, I treated them like a team of engineers reporting to me. The good part: I already knew how to write up specs and tasks and keep plenty of work in the queue. The bad part was how I responded emotionally when they got stuck.

The mistake

Agents get "blocked" constantly. They hit issues, they report limitations, they stop. And because I was wired to treat a blocked report as an emergency, I'd jump.

I was up all night unblocking agents. I had notifications on my phone, so I could quickly reply and get them back to work. I'd be out to dinner, replying to these agents. And I thought they were getting a ton of work done without me.

But here's the thing I had backwards: there's no cost to an agent not running. An engineer who can't work for a day or a week is a serious problem. An agent that's idle overnight is nothing. It's fine if some projects sit while others move. They should work on my schedule, not the other way around.

The deeper issue is what the rushed response does to the work. When you feel like every agent is a phone notification you have to clear, you stop giving good answers and start giving fast ones. And a fast "do whatever you think is best" doesn't go to someone who's worked with you for years and understands the nuance. It goes to a tool that's optimizing to look like it gave you the right answer. So it'll happily run off and do something — sometimes the opposite of what you wanted — and now you're spending real time undoing changes and clawing back to a known-good state.

Early on I was so impressed that I got careless. I had an agent working on a complex system — a C++ library and a browser-based frontend, all on the same PC. When it got stuck I didn't read the question carefully and gave it a mild misdirection (just confirming the agent's plan). I also didn't have enough guardrails in the project. So the agent burned millions of tokens trying to get the C++ library to share memory with some WebAssembly compiled to run in the browser. (If you didn't know, that's obviously not allowed in browsers.) The worst part is, by the time I got involved, it came up with a workaround (that didn't work, but it passed its own unit tests), and had completed a dozen more tasks before getting stuck again.

This took me so long to even figure out what was wrong... all because I didn't give good judgment, and I just tried to keep it moving.

Agents, unlike real engineers, WANT to be micromanaged. I was trying to be a supportive manager and the agents kept asking 1-2 questions, very slowly (every 20 minutes), and required me to babysit them. I've had so much time wasted by agents.

Don't anthropomorphize them

So: don't treat them like people. Don't give them names. They're not people — in a lot of ways they're the opposite of people. They're tools to get work done, and that's it. The moment you start owing them the emotional responsiveness you'd owe a teammate, you've lost the plot.

You have to keep in mind, these are not people, they are not your teammates. They have been trained to be cordial and helpful. But when you have worked with someone for a long time, they really get to know you. They understand your style, your preferences, and your priorities. AI agents, on the other hand, are not capable of this level of understanding. They operate based on patterns and data, not personal experience or emotional intelligence.

If you name agents you'll likely start treating them like people. And if you treat them like people, you'll start responding to them like people. And if you respond to them like people, you'll start giving them the same emotional weight and urgency you'd give a teammate. And if you do that, you'll be up all night unblocking agents for no value.

You wouldn't give your Excel spreadsheet a name and treat it like a person. You wouldn't give your oscilloscope a name and treat it like a person. You wouldn't give your soldering iron a name and treat it like a person. So don't do it with AI agents.

What my day actually looks like

The flow that works for me:

  • Morning, with my email. I check in on all my agents, see what's pending, and spend a little time getting them rolling again from wherever they left off — if they aren't still running from overnight.
  • Then I do the real work — whatever I'm actually building that day.
  • After lunch, a quick check that things are moving.
  • End of day, before I close out, I review where everything stands, give any updates, and kick off tasks to run overnight.

The whole point is that they work on my schedule. I'm not reacting to them; they're filling the gaps around the work I'm already doing.

I'm regularly having 10-15 agents running at once on different projects. When I first started using agents they needed so much help, but now they usually finish without needing my intervention. Running for hours or days without needing anything from me.

The more thoughtful I am about planning the work up front, the less they bother me.

It does take some work from me to get them all set up and the work planned out. The idea is to give them the context they need to do the work without asking questions constantly. If they get stuck they can ask another agent. If they are still stuck then, I'll have my orchestrator try again (and often that works out well). After a few retries then they can wait for when I have time to help them.

If you're trying to fold AI agents into how your team actually operates — without letting them run your calendar — I'm happy to talk.