- Harjot Gill argued that, across the industry, the pull request becomes the planning and review gate. At CodeRabbit, he said, humans review releases and their risk instead.
- He said developers skip human review based on their record in production. A rollback, he said, takes 10 to 15 minutes.
Harjot Gill, founder and CEO of AI code review company CodeRabbit, told The Pragmatic Engineer that human code review is dead, but human review is not. He argued the planning gate is moving to the pull request, the page where a code change is proposed. At CodeRabbit, he said, people review releases and their risk.
What was said
Gill said planning once protected scarce engineering time. AI opened that floodgate: a support ticket, a designer or a background agent can now turn an idea straight into code. So the review gate, he argued, becomes the pull request. A person decides whether to ship the change or close it as slop, then asks about its risk. He said "the human taste is the bottleneck now, human judgment."
At his firm the unit of change is a release. He said "each release is what we are reviewing now, not PRs as a human." Engineers may not read the code, he said, but they understand the risk, and they handle any rollback, which he put at 10 to 15 minutes.
Trust is earned from production history. Some developers are on a review-exempt list. Gill said the company traces each regression to the change that caused it. A careless developer, he said, can leave 100 hours of exposure, when a customer has to point out the fault. After enough strikes, that person returns to human review.
Why it matters
Our reading: the skill being tested is deciding what can go wrong, not typing the fix. Gill's approach lets many changes pass without a person. Scarce attention goes to risky ones, such as a migration on a large table, which he said would crash in production.
A short rollback only limits the loss once someone notices the fault. Gill's own 100-hour example shows that detection, not the rollback, is where the exposure builds up. He also said alerts are limited by the failure scenarios you imagine.
For managers, the question becomes what record lets someone skip review, and who can read a risk label and say no.
The other side
Gill sells AI code review, so he gains if review moves to tools. He described his own firm, not the industry, and called his claim of fewer outages an observation he still had to measure.
The host voiced the opposite worry: quality seems to have dipped across the industry as review has thinned. Gill himself said he has not read a single line of the Rust his agents write.
Training is left open. Gill said CodeRabbit lets interns ship directly to production, but he did not explain how newcomers build the tribal knowledge that spots a crashing migration.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- The Pragmatic Engineer: Software factories and code reviews with Harjot Gill (2026-11-18)





