TL;DR
CodeRabbit founder and CEO Harjot Gill said he shipped more than 5,000 pull requests in a month using AI agents, while the company spends an estimated $250,000 to $275,000 monthly on AI. In a Pragmatic Engineer Podcast interview, he described using agents for infrastructure work and argued that teams should shift human review from individual code changes to plans and releases, backed by testing and staged rollouts.
CodeRabbit founder and CEO Harjot Gill said he shipped more than 5,000 pull requests in one month with AI agents, describing a software-development operation that spends an estimated $250,000 to $275,000 a month on AI. In an interview on The Pragmatic Engineer Podcast, Gill said the agents handle work ranging from maintenance to infrastructure, while his team is changing what human review focuses on.
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Gill described using agents for what he called janitorial work, including fixing flaky tests, improving continuous integration (CI) and refining caching. The source says many pull requests are generated while he is commuting or outside standard work hours. He also steered agents on some user-interface changes. The 5,000-plus figure refers to pull requests he shipped in the month discussed; the source does not break down how many were routine maintenance, product work or infrastructure changes.
Among the larger projects cited was a write-ahead log built in Rust on top of object storage. Gill said an agent worked on it for six weeks and that it is now in production. He also described moving CodeRabbit services to Bun over a weekend and building a replacement CI system after performance problems with the existing one. According to the interview account, a vendor had quoted $13 million a year for a CI system capable of handling the company’s pull-request output; Gill said the agent-built alternative took a week and cost 50% less than that quote, with similar performance.
The episode characterizes CodeRabbit as spending millions of dollars per month on AI; Gill’s personal estimate for his own monthly use is $250,000 to $275,000, with an estimated cost of about $150 per pull request. The source also describes the company as having a 40-person engineering team. Those figures are presented in the interview material, not as independently audited financial disclosures.
Software factories and code reviews
CodeRabbit CEO Harjot Gill says AI agents shipped thousands of pull requests. His approach shifts human attention toward plans, verification and releases.
Where agents do the work
Gill described agents handling maintenance, infrastructure and some product changes.
Clear recurring work
Agents fixed flaky tests, improved continuous integration and refined caching.
Build core systems
Projects included a Rust write-ahead log and a replacement CI system.
Support product changes
Gill also steered agents on some user-interface work.
Shift review to the release
When agents produce many changes, teams need checks that can test work before people approve a release.
People review intended work and guide the agents.
Use simulations, API emulation and development environments to check behavior.
Use feature flags and gradual releases, while monitoring logs and performance.
What the account leaves open
The interview describes Gill’s operating approach, but does not establish its business results.
- Agents built systems Gill says are in production.
- CodeRabbit uses several layers of testing and staged rollouts.
- Gill frames high AI costs as research and development spending.
- How often agent changes cause defects.
- Whether products built this way generate paying customers.
- How much human time release review requires.
A High-Spend Model for AI Development
Gill’s account illustrates a case where AI costs are treated as research and development spending, rather than only as a way to reduce the cost of existing engineering work. The interview’s host frames CodeRabbit as an outlier compared with other developers’ reported AI spending of roughly $200 to $2,000 per developer each month. Those figures are the host’s comparison, not a market-wide survey, but they underline how unusual Gill’s stated spending appears.
The business case, as described in the episode, is that a developer-tools company may use agents to build or improve products it can later sell. That logic may not apply to companies with an established product and a customer base that is not willing to pay more for new capabilities. Gill also argued that startups need more funding as AI development costs rise. Whether the spending produces commercial returns at CodeRabbit, or can be replicated by other companies, is not established by the source.
The workflow also bears on engineering jobs and responsibility. If agents produce more changes than people can review line by line, teams must decide how to check systems and releases without treating generated code as automatically reliable. Gill’s approach emphasizes verification, with human attention directed toward plans and releases rather than every individual pull request.
Products Worth Considering

Samsung 27-Inch Odyssey G5 QHD Gaming Monitor (G50F)

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Code Review to Software Factories
Gill has worked in software and founded infrastructure startups for about 30 years, according to the podcast description. The source says that in 2021 he expected AI to write much of the code. At FluxNinja, he saw code review becoming a bottleneck as GitHub Copilot’s autocomplete features sped up code writing. He initially expected demand for AI code review to come mainly from larger organizations already buying products such as Snyk and SonarSource; he later said the market was broader than he had anticipated.
CodeRabbit sells AI code-review tools, but the interview describes its own internal development as a wider “software factory” model, with agents contributing to systems and product work. Gill’s view is not that all human oversight disappears. He argues that engineers should review the intended plan and completed release, while software-based checks test changes during development.
He described several verification layers: formal methods such as TLA+ and Lean; deterministic simulations that test different orders of concurrent operations; emulation of real service APIs; and development environments where agents can check whether services work. For production changes, he cited feature flags and staged rollouts, with agents monitoring logs and performance as releases expand. The source says TLA+ helped CodeRabbit engineers find race conditions that code review had missed.
“Human code review is dead, long live the human review!”
— Harjot Gill, as quoted in the podcast source
What the Spending Figures Show
The source does not provide a publication date for the episode or independently verify Gill’s monthly spending, pull-request count, project timelines or cost comparisons. It is also unclear whether the quoted monthly AI spending refers to a consistent billing period, which services it includes, or how much was spent specifically on development rather than other uses.
The CI comparison is presented as Gill’s account: the source does not identify the vendor behind the $13 million annual quote, detail the cost calculation for the replacement, or provide comparable performance measurements. Nor does it report revenue or product sales attributable to the agent-built systems. The source fragment ends while describing CodeRabbit’s “review exempt list,” so the criteria for exemption and how human review is enforced are not available.
How CodeRabbit Checks Agent Work
The source does not announce a forthcoming product release, financial disclosure or formal review of CodeRabbit’s results. The next practical test of Gill’s approach is whether the company can sustain its agent-driven output while keeping reliability and production performance within acceptable limits. Its described safeguards include simulations, API emulation, development deployments and gradual production rollouts.
For readers evaluating the model, the missing details are measurable outcomes: how often agent-generated changes cause defects, how much human time release review takes, and whether the high AI bill leads to products customers will pay for. The interview offers Gill’s operating account, but does not resolve those questions.
Key Questions
What did Harjot Gill say he shipped?
Gill said he shipped more than 5,000 pull requests in one month with help from AI agents. The source does not provide a detailed breakdown of their contents.
How much does CodeRabbit spend on AI?
The source describes CodeRabbit’s AI spending as millions of dollars per month and gives Gill’s estimate of $250,000 to $275,000 per month for his personal spending. The figures are interview claims, not audited disclosures.
Does Gill say human code review should end?
Gill argues that engineers may stop reviewing every agent-generated change individually and focus more on plans and releases. He also describes testing, monitoring and staged rollouts as ways to check the work.
What safeguards does CodeRabbit use for agent-built software?
The source describes formal methods, deterministic simulations, emulated APIs and development environments, along with feature flags and progressive production rollouts. It does not provide defect rates or a full account of the company’s review policies.
Source: newsletter.pragmaticengineer.com
