Reviews are in: my Agents have ADD
The best code review I ever gave was four words.
A senior engineer — the kind who’d been writing compilers before I understood what a compiler was — submitted a PR in my first months on the team. I stared at one line for fifteen minutes. Then I typed: “Could we make this simpler?” That was it. He made the change.
Good reviewers don’t review everything. They allocate attention. A senior engineer at a healthtech startup whose name carried so much gravity that everyone auto-approved his PRs. He was capable. He also knew it — and nobody had pushed back on his architectural assumptions in years. Not because he was wrong. Because nobody felt they could. Someone should have. A Staff Engineer at YouTube Gaming with visual instincts that made everyone’s work better, but consistent gaps in test architecture: what they needed was someone focused on test design, not their variable names. A junior engineer whose code was the neatest I’d ever seen — proper abstractions, consistent patterns, no cruft — needed someone to challenge his API design assumptions, not tell him where to put the brackets.
Different people need different review. The skill is knowing which kind of attention to bring.
My agents have ADD.
Not the “scattered and distracted” variant — the opposite. They review everything with the same sustained, undifferentiated focus. Every PR gets the same checklist. No model of what this author specifically gets wrong. No sense of where the real risk lives. They notice everything and see nothing.
The result: a comment about colour vs. color appears in the same list as a security vulnerability. The ratio of signal to noise drops below one. Developers start dismissing reviews entirely, which is worse than no review.
And it gets worse. Authority bias: the confident senior engineer’s speculative architecture sails through unquestioned. Leading the witness: explain your approach in the PR description, and the review confirms you were right. Quota filling: three real issues weren’t enough, so the agent finds two more. Adversarial framing: the reviewer sounds like a prosecutor, and the author gets defensive instead of receptive.
Attention is all you need. The Transformer architecture knows this. The problem is that my agents give equal attention to everything, so it is worth nothing specifically.
The best reviewer I’ve known asked four words at the right moment. My agents file twelve bullet points. I’m still trying to teach them the difference. But boy is that hard to write into a good SKILL.
Related reading:
- Don’t Hate the Agent, Hate the Process — on why agent failures are usually system failures in disguise
Review filed. Priority: low. Item 1 of 12: this sentence could be shorter.