TL;DR
CodeRabbit CEO Harjot Gill told The Pragmatic Engineer that he shipped more than 5,000 pull requests in a month using AI agents, while the company employs 40 engineers and spends millions of dollars monthly on AI, according to the report. He argues that teams need ways to verify agent-written software and says human oversight is shifting from individual code changes toward plans and releases.
CodeRabbit CEO Harjot Gill says he shipped more than 5,000 pull requests in a month with AI agents, in a podcast conversation published by The Pragmatic Engineer. The account offers a detailed example of a software company using agents to build and maintain systems at high speed, while raising questions about AI costs, testing and the role of human code review.
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Gill, whose company sells AI-assisted code review, described using agents for infrastructure and maintenance work as well as product development. Examples in the report include fixing flaky tests, improving continuous integration and caching, and making interface changes. Some agent work ran during his commute or in the evenings, according to the episode summary.
The report says CodeRabbit has a 40-person engineering team and spends millions of dollars per month on AI. Gill estimated that each pull request costs about $150 and put his personal monthly AI spending at roughly $250,000 to $275,000. These figures are his estimates as presented in the report; the source does not provide an independently audited breakdown.
Among the projects Gill described were a Rust write-ahead log built over object storage that ran for six weeks and is now in production, and a migration of CodeRabbit’s services to Bun completed over a weekend. He also said agents helped build a replacement CI system after the company received a $13 million annual quote for an external solution. According to Gill, the in-house system took about a week and cost 50% less than that quote while delivering similar performance.
How Agent Work Changes Review
Gill’s account illustrates how AI development costs can function as research and development for a software company, rather than simply as a way to reduce engineering headcount. The report’s author cautions that CodeRabbit may be an outlier: its tools are closely tied to AI development, and faster experimentation could help it create products to sell. That case may not apply to companies with established products and customers who will not pay more for added capacity.
The speed also makes verification central. Gill said agents need to be able to test and improve their own work, a process he described as “hill climbing.” His team’s methods range from formal models and simulations to test environments and staged production releases. If agents generate changes faster than people can inspect each one, the reliability of those checks can shape how much of the work is safe to ship.
Gill expects human attention to move away from reviewing every agent-generated code change and toward plans and complete releases. That is his view of where software work is heading, not a settled industry-wide outcome. The practical question is whether teams can preserve clear accountability and catch failures as their volume of changes grows.
Products Worth Considering

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptics – Graphite

Bose QuietComfort Ultra Wireless Noise Cancelling Earbuds, 2nd Gen, with 20W 2-Port Wall Charger
As an affiliate, we earn on qualifying purchases.
From Code Review to Software Factories
Gill told the podcast that he anticipated in 2021 that AI would write much of the code. At FluxNinja, he saw code review becoming a bottleneck as GitHub Copilot autocomplete tools helped developers write code faster. He said he initially underestimated demand for AI code review, believing buyers would mainly be larger companies already purchasing code verification products. In his account, demand proved broader.
The episode frames CodeRabbit’s operating model as a software factory: agents do repeated engineering tasks and create larger systems, while engineers guide work and check its results. For verification, Gill cited use of TLA+ to identify race conditions missed in code review, experiments with Lean, and a simulation harness that tests different orders of concurrent operations. The team also emulates external APIs and uses development deployments to check whether services work.
For production changes, Gill described feature flags and progressive rollouts. He said an agent can monitor logs and performance during a rollout and continue expansion when conditions appear stable. The report also notes that his spending is far above the $200 to $2,000 per developer per month that its author hears about elsewhere; those figures are anecdotal comparisons, not a market survey.
“Agents need to be “hill climbing.””
— Harjot Gill, CodeRabbit CEO, as quoted in The Pragmatic Engineer episode summary
Costs and Quality Still Need Evidence
The episode summary does not specify the calendar month in which Gill shipped the 5,000-plus pull requests, nor does it provide a detailed accounting of what counts as a pull request or the full AI expenditure. The spending and performance figures are attributed to Gill, and the report does not include independent verification or a comparison of the in-house CI system’s performance over time.
It is also unclear how much human work was needed to guide, revise or approve the reported changes, and how the company measures defects or long-term reliability across agent-generated projects. Gill’s expectation that people will focus on plans and releases rather than individual changes is a prediction; the episode does not establish that the same workflow will suit other companies.
Testing the Factory in Production
The report is a podcast conversation rather than an announcement of a new product or policy, and it does not name a specific follow-up milestone. The next evidence to watch for is whether CodeRabbit can sustain its reported pace while publishing more detail on costs, reliability and review practices. Gill’s examples point to staged deployment and automated monitoring as continuing parts of the company’s approach, but the source does not provide dates or results for future rollouts.
Key Questions
What is the news about Harjot Gill?
The Pragmatic Engineer published a podcast conversation in which CodeRabbit CEO Harjot Gill described using AI agents to build and maintain software, including shipping more than 5,000 pull requests in a month.
How much does CodeRabbit spend on AI?
The report says the company spends millions of dollars per month. Gill estimated his personal monthly spend at about $250,000 to $275,000; the figures are attributed to him and are not presented as audited accounts.
What does Gill mean by agents “hill climbing”?
He means agents should have ways to test and verify their own work, then use those results to improve it. His examples include simulations, formal methods, development environments and staged production rollouts.
Does Gill think people will stop reviewing software?
No. His argument is that human review may shift from inspecting each agent-generated code change to examining plans and broader releases. That is his view, not a confirmed industry-wide change.
Source: newsletter.pragmaticengineer.com
