The Human-in-the-Loop is Tired
pydantic.dev ● Covered by 3 sources
A Pydantic engineer wrote an essay on the emotional toll of AI-assisted coding. He says the code writes itself but reviewing it feels worse, not better.
Based on reporting by pydantic.dev — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
There's a specific kind of tired going around dev teams right now, and a Pydantic engineer just gave it a name: supervision fatigue. Not burnout from writing too much code, but exhaustion from reviewing endless AI-generated output while trying to hold the actual intent of a project in your head. Pydantic, the company behind popular Python validation and AI-agent tooling, builds the stuff meant to make LLM software reliable. Even so, its own team is struggling with the emotional whiplash of using it daily.
The essay leans on a few blunt anecdotes. Colleague Douwe, who maintains the Pydantic AI framework, described waking up to thirty pull requests generated overnight by other people's AI agents, each needing a fast human judgment call. Another engineer, Marcelo, joked about running five Claude Code sessions at once because you'll always be too busy giving feedback to notice one has frozen. Funny, until you realize the joke describes an actual workflow. The number of things you can start has exploded. The number you can finish hasn't moved, because finishing still runs through one unparallelizable bottleneck: a human brain.
The piece borrows a machine-learning term, reward function, to explain why this feels bad even when it's productive. Writing code by hand used to hand out small dopamine hits constantly: solving a tricky bit of logic, watching a build pass, feeling in control. AI assistance automates a lot of that satisfying part and replaces it with the tedious part, reviewing and correcting. Output goes up. Satisfaction goes down. And there's a social cost too: no one's on the other end of that pull request actually learning anything, because it came out of what Douwe calls
Read more about this at: pydantic.dev
