Matthew Boston

Review the outcome, not the output

March 26, 2026

Key takeaway: When an agent writes the code, ask whether the change met its goal, and let linters and tests check the lines.

When an agent writes the code, review the outcome, not the output. The output is the diff. The outcome is whether the change solved the problem, and that’s the question only a person can answer.

Line-by-line review made sense when a person chose every line. An agent didn’t pick that variable name because it was tired. It followed a pattern. Nitpicking its syntax is like copy-editing a compiler’s output. That doesn’t mean quality stops mattering. You just check it differently.

An agent can produce twenty implementations that all pass the tests, and most are fine. You can even ask for several on purpose and keep the best. Whether the feature should exist, solves the right problem, fits the architecture, and helps the user is judgment. That part is yours.

Outcome review is just as rigorous. It looks at different things:

  • Intent. Does this match what was asked? Agents follow instructions literally and don’t question whether the instructions were right.
  • System impact. Does it add coupling that hurts later, or cross boundaries it shouldn’t?
  • User impact. A technically correct change that makes the product worse is still a bad change.
  • Tests. Check that they test the thing that matters, not just that they exist. Agents will happily write tests that pass without testing anything.

Letting go of the diff is only safe when linters, type checks, and tests catch the output-level mistakes before it reaches you. For trivial changes like a patch-level dependency bump, you may not need a person in the loop at all.