LGTM is not a review
Teams approve agent pull requests nobody read, and the green check has turned into theater. An approval should mean you read the change and you're on the hook for it, and a pull request too big to read should get a no.
Say someone on your team opens a pull request with 1,200 changed lines across 30 files. Ninety seconds later it has a green check and a comment that says “LGTM.” Branch protection is satisfied, the merge button turns green, and the change is in production by lunch.
Nobody read it, and everyone involved knows that. It was approved because the process asks for an approval and somebody supplied one.
Approving without reading was a bad habit back when people wrote all the code. Now an agent can open a pull request like that every few minutes, and the “author” often didn’t read it closely either. On a lot of teams the rubber stamp has gone from an occasional shortcut to how code gets merged. The green check looks the same as it always did, but it tells you much less than it used to.
What the button actually says
GitHub’s whole definition of Approve is that it “signals that the changes are ready to merge.” It says nothing about reading or understanding the change. The same page says required reviews “help protect important branches and reduce accidental merges,” which is true only as long as the person approving does something before they click.
So the button doesn’t vouch for anything by itself. An approval means whatever your team has decided it means, explicitly or by habit, and on a lot of teams habit has settled on “I saw this go by.”
What you owe as a reviewer
Google is more direct about this with its own engineers: “a reviewer has ownership and responsibility over the code they are reviewing.” Its guide tells reviewers to look at every line they’re assigned and to “be sure that you understand what all the code is doing.”
The guide makes one exception, and it matters more now than it used to. Reviewers can sometimes skim data files, generated code and large data structures, but they shouldn’t skim a human-written class or function and assume what’s inside is fine.
I suspect a lot of tired reviewers are filing agent code under “generated,” and that’s a mistake. As I read it, the exception is for output nobody maintains by hand, like code generated from a schema, which gets regenerated instead of edited. Agent code is code your team will read, edit, debug and get paged for, so it should get the same line-by-line review as code a person wrote.
“Too big to read” is a complete review
The usual excuse for an LGTM is that the pull request was too big to review properly, so it got reviewed improperly. That reasoning has it backwards.
The guide also lets a reviewer say no. Reviewers can reject a change outright for being too large, and if the code is too hard to read, the guide’s advice is to tell the author and wait for them to clarify it before reviewing.
Agents make saying no cheaper. Splitting a pull request used to cost the author an afternoon of untangling commits, and reviewers felt bad asking for it. Now an agent can split its own pull request into a refactor and a behavior change in a few minutes, so asking for one costs the author very little. On a 1,200-line diff, “this is too big for me to read, please split it” is the most useful comment a reviewer can leave, and the approval should wait until the reviewer gets something they can read.
What an approval should mean
Since the button doesn’t define an approval, I think each team should define its own and write it down. Here is the definition I’d use. Before you click Approve, you should be able to say yes to all of these:
- You read every line, or you said in the review which parts you didn’t read and who should.
- You can say what the change does differently when it runs: which routes return something new, which tables get written, what happens when a call it makes fails. (I wrote about putting that list at the top of the pull request, generated from the code.)
- You’d be fine being the one who gets paged when it breaks.
- You could explain it in the incident review without opening an agent to ask what it does.
The bar is high on purpose, but partial approvals are fine as long as they say what they cover. “Reviewed the migration and the API changes, didn’t look at the UI” is more useful than a bare LGTM, because the next reviewer knows where to look.
If nobody on the team can clear that bar for a change, the change isn’t ready to merge. That’s a finding about the change, and a review that turns it up has done its job.
Why LGTM keeps winning
LGTM keeps winning because of what teams reward. Review work is mostly invisible. Shipping a feature shows up on the roadmap and in your promotion packet, but an afternoon spent reading somebody else’s diff and catching the bug that would’ve double-charged customers shows up as a comment thread. When review time competes with work that gets counted, review loses, and it loses fastest on the big pull requests that need it most.
There’s also a social cost to saying no. Asking a colleague to split their pull request can feel like an accusation. It shouldn’t when their agent wrote the pull request, but the request still lands on the colleague.
If you want approvals to mean something, the team has to pay for that:
- Count review as work. Bring it up in one-on-ones and in promotion cases, and look at what reviewers caught, not only at what authors shipped.
- Make “too big, please split” a normal comment that senior people leave first, often, and without apologizing.
- When an incident traces back to a change, look at the approval too. Could the reviewer have explained the failing path before it merged? Treat that as a question about the process, and change the process when the answer’s no.
A green check should say that somebody is on the hook for the change. If your team lets it mean less than that, review produces entries in the audit log and little else, while agents open pull requests faster than anyone can read them.