Skip to content
Innovation & Growth Talking point

Gill: small teams shipping with AI agents now need big-company release tooling

CodeRabbit's CEO argues staged rollouts and merge queues, once big-company tooling, are now needed by small teams.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
Gill: small teams shipping with AI agents now need big-company release tooling
In brief
  • Harjot Gill of CodeRabbit argued that teams of 30 or 40 people shipping with AI agents now need release tooling once built for Facebook and Netflix.
  • Gill's stated goal is not zero outages but the shortest possible time a bad change stays exposed in production. He runs an AI code review business, which is worth weighing.

Harjot Gill, founder and CEO of CodeRabbit, argued that the release tooling built for Facebook and Netflix is now needed by teams of 30 or 40 people. He made the case on The Pragmatic Engineer, in an episode published on 18 November 2026. His reasoning: once AI agents let a small team ship at giant pace, it needs giant-style safeguards.

What was said

Gill said CodeRabbit ships code every 30 to 45 minutes. New features go out behind a feature flag, a switch that turns a change on for only some users. A release may first reach about 10% of open-source users. An agent watches performance and logs, then widens the rollout once things look stable.

The host noted that staged rollouts were long a Facebook specialty. They seemed to need hundreds of millions of users and costly machine learning. The host put CodeRabbit at perhaps 40 engineers. Gill replied that the tooling big firms needed, from merge queues to testing in production and managing risk, is what small teams will now need if they use agents well. A merge queue is the line that checks changes before they join the main code. He said flaky tests keep resetting it, so agents work to remove them.

On outages, Gill said total prevention is not the aim. The goal, in his words, is making sure "the exposure time of something bad happening in production is as little as possible".

Separately, CodeRabbit uses formal methods, which means writing a precise mathematical model of how a system should behave and checking against it. It modelled a Slack bot's distributed system in TLA plus before building it, and it is adding Lean theorems for things like a billing service. Gill described this as CodeRabbit's own practice. He did not say every small team must copy it.

Why it matters

Our reading: AI raises how much change a small team can push out, and the weaknesses in its release process grow with it. A team that ships every half hour cannot have a person read each change first. Gill said he has not read a single line of the Rust his agents wrote.

Limiting exposure can stand in for that review because of how it works. A small first audience means a faulty change harms few people. Fast observation means it is caught and pulled before it spreads.

A manager or buyer can ask an engineering lead three plain questions. Can a release reach a small share of users first? How many people can one bad release hit? How long does it take to notice and undo a fault?

The other side

Gill runs an AI code review company, so he gains if teams believe AI creates new needs. That is worth weighing when reading his claim. His view that 30 or 40 people can ship as much as Facebook came with a condition: that they use agents in the right way. It is his view, not a measured result.

The host also made clear that this machinery was hard to build, and Gill gave no cost for a small team to match it. His own goal accepts that some bad changes will still reach production. The aim is to limit the damage, not to prevent it.

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