AI Tools and the Productivity Illusion

Gergely Orosz nailed something I've been chewing on:

The more I use AI tools, the more I have to admit that I'm not that much more productive... I simply FEEL that much more productive. In reality, the context switching of kicking several things off wipes out my perceived productivity gains.

Yeah. The gap between feeling productive and being productive is sneaky, and I keep falling into it.

The feeling is real

Kick off an agent to scaffold a feature, another to write tests, start drafting a PR description. Five minutes in and you've got all this stuff moving. You're not staring at a blank file. You're orchestrating.

It's intoxicating. And it's easy to confuse that momentum with actual output.

But you pay for it

So you've kicked off three things. Now you have three things to babysit. Each one comes back needing review, correction, or another prompt. You're not writing code anymore, you're managing code. And switching between agent output A and agent output B costs you the same cognitive tax as any other context switch.

The classic context-switching penalty doesn't disappear just because AI handles the execution. It might actually get worse. The barrier to starting new work is so low that you end up juggling more threads than you ever would have created by hand.

I've been here before

When I moved from engineering into management, I hit the same trap. Just with people instead of agents.

As an IC, I had one stream of work. Mine. I could hold the full context in my head and push it forward. Then I became a manager and could delegate. Obviously, if I hand things off to four people, I get four times as much done. Right?

No. I got four concurrent threads to stay on top of. Four PRs to review, four sets of context to hold, four conversations to track. Delegation removed the execution bottleneck and replaced it with a coordination bottleneck. The cognitive load of staying across all of it often left me feeling busier than when I was just writing the code myself.

The lesson took a while to land: be deliberate about how many things are in flight. Not "how many things can my team work on" but "how many things can I actually support right now." Always fewer than I thought.

Same lesson, different costume

AI tools compress that same dynamic into one person's workflow. You're the IC and the manager. You delegate to agents like you'd delegate to a team, but you're still the one reviewing, course-correcting, and stitching the output together. Still one brain.

The times I've gotten real gains from AI share a pattern: one thing at a time. Use the AI to move faster through a single task, not to spray work across five. Same principle that made me a better manager. Don't parallelize past the point where you can keep quality up.

The bottleneck isn't execution anymore. It's impulse control. Starting something new is nearly free, whether that's assigning a ticket or firing off a prompt. The hard part is not starting the next thing before you've finished the current one.

I don't think the people who get the most out of these tools are the ones running ten agents in parallel. They're the ones who figured out, probably the hard way, that how many things you can start was never the constraint. It's how many you can finish well.