Agents Write My Code. I Still Read the Diffs.

I can hand off the typing. I can’t hand off knowing what’s in prod with my name on it.

It’s 4 PM on a Friday. The incident channel is on fire, the API is throwing 500s, and the Datadog graph is a straight cliff down. I’ve been here before; I pull up recent GitHub PRs and start reading commits I’ve never actually seen, because my agent wrote them, committed them, opened the PR, and I merged it earlier in the day without really looking. This is my first real pass through the code. Just before the weekend, under fire.

I’m not writing this to complain about agents. They write most of my code now. They commit, they open PRs, they write better commit messages than I ever bothered to. That fight is over.

But one habit died in the transition, and I’ve had to consciously bring it back: reading what ships. Not reviewing. Reading. Different thing. It’s the one thing I don’t let the agent take from me.

Photo by Fotis Fotopoulos on Unsplash

The diff is the one artifact in the whole agentic loop that’s actually made for humans. I don’t audit the 400 tool calls. I don’t trace the reasoning. I don’t reconstruct the plan. The diff is what changed between prod yesterday and prod tomorrow, small enough to skim.

Here’s what I get out of it.

Incident context, ahead of time. Every incident starts with “what changed?” If I’ve read the diffs going in, I have a rough map before the incident channel lights up. “Didn’t we touch retry logic in the payments client yesterday?” is a 20-minute incident for me. For someone reading that diff for the first time, it’s a 3-hour one.

Accountability that actually holds. The agent doesn’t show up to the prod incident call. I do. “Claude wrote it” lands about as well as “the intern pushed it.” Nobody in the room cares. My name is on the PR, so my eyes should have been on the change. That’s the deal I made when I hit merge.

A mental model that doesn’t rot. Every skipped diff is a small withdrawal from my understanding of the system. Do it enough and eventually I can’t reason about my own codebase. I can only ask the agent about it. That stops being leverage and starts being dependency, and I’ve watched it happen to people around me.

What reading actually means here, because I’m not line-reviewing like it’s 2019: I skim the file list first. Just knowing what got touched is signal. “Why is a copy change modifying auth middleware?” has caught more real bugs for me than any linter or pre-commit hook.

I actually read the parts near the blast radius. Auth, payments, migrations, config, GitHub Actions workflows, ArgoCD manifests, LaunchDarkly flag changes. Boring files cause loud incidents. I’ve never regretted the extra minute on a workflow or manifest diff.

I sanity-check the shape. If I asked for a small fix and the diff is +1,400 / -300, something is going on. I don’t need to read it. I need to reject it and ask what happened.

And I watch for stowaways. Agents like to do drive-by refactors, opportunistic dependency bumps, “while I was in here” cleanups. Every one of those is an unreviewed change hiding inside a reviewed one. I had a Claude-authored PR once that quietly rewrote a caching layer while fixing a typo. The typo fix was fine. The caching change caused a Sev2 four days later. I’ve been slower to trust small PRs since.

Five minutes. That’s usually all it costs me.

Letting agents write, commit, and ship is outsourcing labor. Good. I do more of it every month. Skipping the diff is outsourcing knowledge, and knowledge is what I get billed for at 2am, with interest. I can hand off the typing. I can’t hand off knowing what’s in prod with my name on it.

Merge the PR. Thank the agent. Read the diff first. That’s the order I try to hold to.