Simon Willison nailed it. In a conversation with Lenny Rachitsky doing the rounds this week, he said using coding agents well takes every inch of his 25 years of experience as a software engineer, and by 11am he’s wiped out for the day. A 48-second clip has cleared 1.1 million views. That ratio tells you something.
I’m not a professional developer. I’m a dabbler, an enthusiastic amateur who builds things for fun and curiosity. But even at my level, the fatigue is real. The bottleneck is not the code generation. It hasn’t been for months. The bottleneck is me. My ability to read, assess, test and redirect the output. The agents will happily churn through tasks faster than any human can meaningfully evaluate the results.
Willison calls this agentic engineering and he’s right that it’s a new discipline. You’re not writing code. You’re air traffic control. Scanning diffs, verifying behaviour, catching bugs, making architectural calls on the fly. If Willison is cooked by 11am with 25 years of engineering behind him, the rest of us don’t stand a chance at sustained focus.
And yet. This stuff is addictive.
I catch myself on the couch at 10pm, iPad propped up, half watching TV, half prompting an agent to enter plan mode to explore new functionality. It’s the worst possible environment for the kind of sharp attention agentic coding demands, and I keep doing it anyway. The dopamine hit of watching an agent build something in minutes that would have taken me days is genuinely hard to put down. Honestly it helps me build things I never could before. Not even attempted. Over this weekend it’s been my first iOS app.
Traditional coding had natural pauses. Waiting for a build. Staring at a function while the kettle boils. Those pauses weren’t wasted time. They were cognitive recovery. Agentic coding strips all of that out. The agent doesn’t need a break. It just keeps going. And it expects you to keep up.
The result is a new kind of tired. Not the satisfying exhaustion of building something from scratch. More like the drained feeling after a day of back-to-back decisions. You got a lot done. You can’t do any more. And you probably let a few things through that you shouldn’t have.
The tools will keep getting faster. The human in the loop has a fixed cognitive bandwidth. The people who figure out sustainable pacing for this work will outlast those of us doom-coding from the couch at midnight.
The limiting factor isn’t the AI. It’s the humans.
Sources:
- The cognitive impact of coding agents (simonwillison.net)
- An AI state of the union, Lenny’s Podcast (lennysnewsletter.com)