Ball posts the push-to-main pipeline behind his no-PR team
After a week of arguing that code review adds nothing, he answered the how with a build and deploy setup.
Thorsten Ball closed out his argument against pull requests by publishing the pipeline his team actually runs. Everyone pushes to main. Buildkite runs the tests and builds images, GitHub Actions then runs infrastructure security checks and kicks off deployments, and a failed check blocks the deploy and pings the commit author in Slack. CI keeps running on later pushes while the break is being fixed, and a later fix can turn it green again and re-enable deployment. Staging then prod usually takes about 10 minutes. Durable agent infrastructure deploys on a two hour cadence, CLI releases go out every four hours.
That is the answer to the obvious question his earlier posts left open. His original case was that human diff review no longer earns its place, because you can spawn Opus 5.5, Fable and Astra and have each of them check four times. He also argued that CI originally meant the opposite of long-lived branches, and that having agents run the suite, then running it again on a branch, then again on merge, does not make sense.
Where the pushback landed
The sharpest objection came from Benjamin Glickenhaus, who said the part he was missing is how work in an isolated environment gets back into the shared codebase, and how teammates arrive at consensus on the APIs and contracts of shared tooling. He described his own habit of telling a teammate to have their agent build on top of a change and report what works, with the feedback folded back in. His framing was that teams need a continuous consensus protocol, since pull requests are discrete moments of consensus building.
Ball's reply to the wider skepticism was not technical. He said it sounds like a lot of people do not trust their colleagues to work like this was their own project. He also claimed a team can ship fast, ship a lot and build a beloved product with no code reviews and no PRs, without yoloing to main.
For working engineers, the useful detail is the shape of the safety net rather than the slogan. Nothing in the pipeline gates a change before it lands. Everything gates the deploy, with the blast radius controlled by how fast a broken main can be fixed and how loudly the system tells the person who broke it. Whether that substitutes for a colleague agreeing with you about an interface is exactly the point still in dispute.
Sure sounds like a lot of people don't trust their colleagues to work like this was their own project.
It feels like we need a more continuous consensus protocol (whereas PRs are discrete moments of consensus building)
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.