If You Stopped Using PRs in the Age of AI, Your PRs Aren't High-Tech Enough

PRs Are Not the Slow Part

I recently put together a CI/CD setup for some of my clients, and it reminded me how easy this has become.

Not easy as in "there is no work." Easy as in: all the annoying pieces that used to make a good pipeline feel like infrastructure for a much bigger team can now be assembled, debugged, and improved with an AI agent sitting beside you.

Then the agent can become part of the pipeline.

So when I hear that pull requests are unnecessary in the age of AI, I think the opposite. If AI made you stop using PRs, your PRs aren't high-tech enough.

The Old Pull Request

The bad version of a PR is ceremony. Someone opens a diff, writes "please review," waits for another person to find time, and hopes that person notices whatever the author missed.

I understand wanting to delete that process. It is slow, inconsistent, and expensive. But the problem is not the pull request. The problem is that almost nothing useful happens when you open it.

A modern PR should kick off the work immediately.

• Build the project.
• Run the tests, types, linters, and security checks.
• Prepare whatever preview or deploy artifact the project needs.
• Have AI read the diff in the context of the repository.
• Flag suspicious logic, missing tests, accidental coupling, and changes that do not match the stated intent.
• Summarize the change so a human can spend attention on the decisions that matter.

By the time a person opens the PR, the cheap questions should already have answers.

AI Can Review the PR

The same class of tool that wrote the code can review it. Better yet, use a separate model or a fresh context so the reviewer is not merely defending its own choices.

An AI reviewer can compare the implementation with the PR description, trace changed call sites, look for unhandled edge cases, ask why a test is missing, and notice that a supposedly local change reached across three modules. It can do that on every update, without making a teammate reread the entire diff.

That does not make the AI correct. It makes review cheap enough to run by default.

Traditional checks still provide the hard evidence: the code compiles, the tests pass, the types line up, the image builds. AI adds a different kind of pressure. It can challenge the assumptions around the change and explain where the evidence looks thin. One model generates; another system judges. I have written about this as a generator-and-discriminator loop. A good CI/CD pipeline is that loop made automatic.

Humans Move Up the Stack

This is not an argument for removing human review. It is an argument for not spending human review on work a machine can do first.

AI can tell you that a change introduced a new dependency. A human knows whether that dependency is acceptable for this client. AI can identify an undocumented assumption. A human knows whether the assumption matches the business. AI can point at code that looks hard to maintain. A teammate can say whether it will be hard to maintain here.

That is the useful human part of review: not manually pretending to be a compiler, but deciding whether the change belongs in the system. And the PR has an even broader job: it organizes the change for every stakeholder around it. Automation clears the mechanical work out of their way.

The PR Is the Control Surface

AI makes code cheaper to produce. That increases the value of a clear checkpoint before the code lands.

The pull request is where you can collect the diff, the intent, the test results, the preview, the deployment plan, and the independent critique in one place. It gives the team, and the next AI agent, a durable record of what changed and why it was allowed through.

This is what I mean when I ask an agent to make something durable. A useful finding should not disappear with the session that produced it. Put it into a test, a check, documentation, or the review record so the next person, and the next agent, inherits it automatically.

The client setup that inspired this post did not feel good because it had a particularly clever YAML file. It felt good because opening a PR started a useful process. The pipeline gathered evidence. AI supplied another set of eyes. The human reviewer started with context instead of a blank diff.

That is what CI/CD should feel like now.

If your pull requests are only a waiting room for human approval, modernize them. Don't throw them away.

Header photo by Christopher Burns on Unsplash.