The Diff Is Where AI-Assisted Development Becomes Engineering

An agent will tell you that it fixed the retry logic, added tests, and preserved existing behavior. That is a claim. It is not evidence. The evidence sits in the repository, and there is one command that shows it.

git diff

We keep coming back to this because it keeps mattering. AI has made it dramatically easier to produce a change but it has not made it easier to know whether that change is the one you actually wanted.

Say an agent reports that it updated retry behavior and added a shared helper. However, the changed files tell a different story. Instead of two fixes, AI edited six files including a configuration change and a dependency bump you did not ask for. Perhaps all six changes are justified but until you actually look, you won’t know that. You only have the AI’s word.

A one-line change can have a bigger impact than a large edit. cache.invalidate(id) becoming await cache.invalidate(id) barely registers as a diff, and it raises the question of whether an invalidation returning a promise changes request latency, and what should happen if it fails. Size and risk are measuring different things, and it is easy to let one stand in for the other.

Large diffs are not a problem to eliminate but they are a cost to plan for. A model can generate twenty-eight changed files in a few minutes. The ability to actually understand twenty-eight file changes did not get faster to match. Breaking a large change into smaller, reviewable steps does not slow the work down as much as it might seem to. It just moves the cost to where it is cheaper to pay.

We built the Crosscheck feature around a fairly simple observation: a model reviewing its own commit is not really review, it is confirmation. Choosing a second model, ideally from a different model family, to look at a commit tends to catch things a single model working alone would not. The review appears in the conversation. What happens with it is still your decision.

The diff has always been the source of truth. Summaries, explanations, and confidence scores are just descriptions of the work, and reading a description is never a substitute for inspecting the code itself.