Glonce

Simon Willison: line by line review isn't the best way to verify AI code

agentsThe Glonce editorsSimon Willison🔥50Glonce rating50Impact35Novelty40Relevance85Surprise30Players45Credibility65

In a short post published on 22 August 2026, developer Simon Willison writes on his blog that the key skill for making productive use of coding agents is being able to confidently instruct them on how to make changes and then confidently verify that those changes were applied correctly. He says reviewing every line of code an agent has written is one way to do that, but not the only way, and states plainly that eyeballing every line of code has never been the most effective way to validate a change to a piece of software. The post argues the point rather than demonstrating it: it does not name or describe what the alternative verification methods actually are, so readers are left with the claim itself rather than a concrete technique to apply.

Key facts

  • Simon Willison published the post, titled 'More than just code review', on 22 August 2026.
  • He argues the key skill for productive use of coding agents is confidently instructing them and then confidently verifying the changes were applied correctly.
  • He says line-by-line review of agent-written code is one valid way to verify changes, but not the only one.
  • He states that eyeballing every line of code has never been the most effective way to validate a change to software.
  • The post does not specify what the alternative verification methods are.

Why it matters

As AI coding agents get used for more of the actual writing of software, how a developer confirms the result is correct becomes a bottleneck of its own. Willison, who writes regularly about AI and developer tooling, is pushing back on the default assumption that thorough verification means reading the agent's output line by line, arguing that goal can be reached other ways.

Who it affects

Developers and teams who use AI coding agents and are deciding how much manual code review to apply to what those agents produce.

How to use it

The post offers a stance, not a method: it does not lay out what the alternative verification techniques look like. A reader takes from it a prompt to question line-by-line review as the default, not a specific procedure to adopt in its place.

How solid is it

This is a short opinion post on the author's personal blog, not a study or a piece backed by data or citations. Its weight rests on Willison's standing as a commentator who writes frequently about coding agents and LLM tooling, not on evidence presented in the post itself.

Risks and caveats

Without a described alternative, the argument could be read as license to skip verification altogether rather than to substitute a different, equally rigorous check. The post is brief and does not address cases where line-by-line review may still be the safer choice, such as security-sensitive or hard-to-test code.

“The key skill required to make productive use of coding agents is being able to confidently instruct them on how to make changes and then confidently verify that those changes have been applied in the correct way.”

— Simon Willison