- Harjot Gill said he ships agent-written Rust he has not read, and relies on automated checks and production monitoring instead.
- A host raised that defect rates may not fall as volume rises. The excerpts give no measured defect figures either way.
Harjot Gill, founder and CEO of CodeRabbit, an AI code review company, described a new way of working on The Pragmatic Engineer. His agents write the code and he does not read it. He said the answer is not more reading but better checks, and shipping speed raises the stakes.
What was said
Gill said his agents write Rust, a programming language, and that he has no idea how it works. One agent ran for six weeks building a Git write-ahead log, a record of changes kept on object storage. It now runs in production and powers knowledge bases at CodeRabbit. In his words: "I have not read a single line of it."
Instead of reading, he said, the agents check their own work in loops. The company uses formal methods, which are mathematical models of how a system should behave. It also uses simulations and emulators of cloud services. Code ships every 30 to 45 minutes, often behind feature flags with gradual rollouts.
Volume is high. The host cited more than 5,000 pull requests last month, and Gill confirmed it. Gill said: "you wouldn't believe everything just breaks when you start shipping at that scale." He said the goal is not to stop every outage but to keep the time bad code stays live as short as possible.
He also described a driving-licence system. Developers whose changes cause regressions lose the right to ship without human review. He said careless prompting can mean sometimes 100 hours of exposure before a customer reports the fault.
Why it matters
Our reading: Gill is describing what replaces a human reader once code arrives faster than anyone can read it. In his account, three things stand in. They are automated verification, agents that watch production, and review rights earned through a clean record.
He added that areas of the code where no agent watched production had longer exposure to bad code. For anyone who buys or runs software built this way, the useful question is what checks the supplier has, not who read the code.
The other side
The strongest limit is that none of this is measured in the excerpts. A host remarked that defect rates may not be falling, and that more defects ship because volume has risen. Gill answered "Yep" and said teams must stay on top of them. So the risk grew with the volume, on the host's framing, and Gill did not dispute it.
Nobody in these excerpts gave defect rates from before or after agents. Everything here is Gill's own account. CodeRabbit also sells AI code review, so he has a stake in the idea that verification can replace reading.
Gill also said human taste and judgment are now the bottleneck. He said humans still review some higher-order abstraction. The conversation did not settle how far that thin layer of human review can stretch.
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)





