Pydantic essay: the human in the loop is tired

Supervision fatigue

Pydantic's Laura Summers has published an essay, “The Human-in-the-Loop is Tired”, arguing that programming with large language models is genuinely useful and genuinely destabilising at once, and that ignoring the second half of that sentence will burn developers out. Pydantic builds data-validation tooling, the Pydantic AI agent framework and Logfire.

The core of the piece is supervision fatigue. Her colleague Douwe, who maintains Pydantic AI, described waking up to roughly thirty pull requests every morning, each opened overnight by somebody's AI, and having to make snap judgements on all of them. Delegating the review to a model was tempting, but, as he put it, “at that point, what am I still doing here?”

Summers describes spending close to two full days writing a plan for a model to execute — clarifying and re-specifying — only for it to port a React hook into the wrong file or invent components that do not exist. Those are errors of coherence, not capability: a model can produce plausible code without holding a coherent intent across a complex change.

The reward function problem

She names the mechanism the human reward function problem. A reward function tells an agent what good looks like; hand-written code was never easy, but it was full of small rewards — solving a problem in your head, watching it compile, feeling in control. LLM-assisted work automates much of that and replaces it with review and supervision: the satisfying part shrinks, the exhausting part grows.

Intensity rises too. Summers cites a Berkeley Haas study on how AI intensifies rather than reduces work, and recalls prompting until nearly 2am because a plan felt almost right. A colleague joked about opening five coding-agent sessions at once: the number of things you can start has grown sharply; the number you can thoughtfully finish has not, because that still needs one brain.

A precedent, and what survives

Summers compares the moment to the 2009 shift from fixed-width layouts to responsive design. Designers hated losing pixel-level control, but the craft did not die: proportion, hierarchy, systems thinking and designing for uncertainty became more relevant. Today's shift is faster and carries more existential dread, yet the pattern holds.

When plausible code is cheap, the distinguishing markers become taste, nuance, mature architectural opinions and contrarian calls earned through real expertise. New practices are appearing too — pre-mortems, where a fresh session must assume a plan failed catastrophically and explain why. Her conclusion is measured: the bottleneck was never the code but human attention and engineering judgement. The humans are still in the loop — just tired.