TL;DR
In an interview with The Pragmatic Engineer, CodeRabbit CEO Harjot Gill described a workflow in which AI agents produce thousands of pull requests and help build internal software systems. He said the company spends millions of dollars a month on AI, while stressing that agent-generated work needs layered testing and human review of releases.
CodeRabbit CEO Harjot Gill says he shipped more than 5,000 pull requests in one month with help from AI agents, offering an account of how his company is reorganizing software development around automated code production. In an interview published by The Pragmatic Engineer, Gill also described spending roughly $250,000 to $275,000 a month on AI and using multiple layers of testing to check agent-generated changes.
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, founder and CEO of CodeRabbit, discussed the approach in a podcast interview about software factories, code reviews and the rising cost of AI-assisted development. The publication says CodeRabbit has 40 engineers and spends millions of dollars per month on AI. Gill estimated that an individual pull request costs him about $150, while describing his personal monthly AI spending as around $250,000 to $275,000. Those figures are his estimates, not independently audited costs.
Much of the reported pull-request volume concerns internal maintenance work, or what Gill called “janitorial work.” He said agents address tasks such as fixing flaky tests, improving continuous integration and refining caching. Some work runs during his commute or overnight. Larger projects described in the interview include a Rust write-ahead log built on object storage, which Gill said ran in production after an agent worked on it for six weeks; a weekend migration of services to Bun; and a replacement CI system.
Gill said the company chose to build that CI system after receiving a $13 million annual quote for a service that could handle its pull-request output. According to the interview, the agent-assisted replacement took about a week and cost 50% less than the quoted option, with similar performance. The source does not provide an independent comparison or further cost details.
The Economics of Agent-Driven Development
Gill’s account is relevant to engineering teams weighing whether the costs of AI tools can be justified by faster development or new products. His argument is that, for a developer-tools company, heavy AI use can function as research and development when it helps create tools the business can sell. That rationale may not apply to companies with a fixed product and customer base that cannot readily pay more for added development capacity.
The spending also sits far above the range cited by the interviewer: roughly $200 to $2,000 per developer per month in other AI spending the publication’s editor said they hear about. That comparison is an anecdotal range from the interviewer, not a standardized industry survey. Gill’s case is presented as an outlier, and the interview does not establish that comparable spending would pay off at other companies.
The reported workflow shifts engineering attention away from inspecting every line of generated code and toward testing, planning and reviewing releases. If agents can produce changes quickly, organizations still need ways to detect defects, manage rollouts and decide which work is safe to ship.
Products Worth Considering

Jackery Explorer 1000 v2 Portable Power Station, 1024Wh LiFePO4 Battery

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptics – Graphite

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten
As an affiliate, we earn on qualifying purchases.
From Code Completion to Software Factories
The interview traces Gill’s interest in AI-assisted software back to 2021, when he expected AI to write much of the code. While working at FluxNinja, he saw code review becoming a bottleneck as GitHub Copilot autocomplete sped up parts of code writing. He said he initially underestimated the market for AI code review, expecting demand mainly from larger companies that already bought verification products such as Snyk and SonarSource. He now sees potential customers across a broader range of companies.
At CodeRabbit, the idea of a software factory means more than asking an agent to write a feature. Gill described agents working on infrastructure and maintenance, alongside checks intended to verify their output. Those checks include formal methods such as TLA+ and Lean, deterministic simulations of concurrent operations, emulated external APIs, development environments and staged production rollouts using feature flags.
Gill said a team using TLA+ found race conditions that code review had missed, and that the company built a simulation harness to test different interleavings of concurrent operations. He also described using agents to observe logs and performance during gradual releases. These are accounts of the company’s practices from the interview, rather than a third-party evaluation of their effectiveness.
““hill climbing””
— Harjot Gill, CodeRabbit CEO, as quoted in The Pragmatic Engineer interview
What the Cost Figures Establish
The spending and productivity figures in the interview come from Gill’s account; the source does not provide independently audited records of monthly AI costs, pull-request counts or the $150 estimate per pull request. It also does not spell out how the reported 5,000 pull requests were divided between maintenance, product development and other work, or how many were merged and used.
The CI comparison is likewise reported without a detailed cost breakdown, testing methodology or independent performance data. The interview describes the new system as costing 50% less than the $13 million annual quote and performing similarly, but it does not establish whether that result can be replicated elsewhere. More broadly, it remains unclear how the company’s AI spending translates into revenue or long-term returns.
How CodeRabbit Checks Agent Work
The interview does not announce a new product launch, financial filing or scheduled milestone. The next developments to watch are whether CodeRabbit continues to expand its agent-assisted engineering practices and how it reports on the cost, reliability and commercial results of those practices.
Gill’s described approach points to ongoing investment in verification: formal models, simulations, API emulation, development deployments and staged releases. For readers evaluating similar systems, the central open test is whether these checks catch failures consistently as agent output grows—and whether the resulting products justify the added AI expense.
Key Questions
What is the news about Harjot Gill?
The Pragmatic Engineer published a podcast interview in which the CodeRabbit CEO described using AI agents to help produce more than 5,000 pull requests in one month and explained the company’s approach to testing their work.
How much does CodeRabbit spend on AI?
Gill described monthly AI spending in the millions of dollars for the company and estimated his personal monthly spending at about $250,000 to $275,000. These are figures reported in the interview, not independently audited totals.
What kinds of work did the agents do?
Gill said agents handled maintenance tasks such as fixing flaky tests and improving CI and caching. He also described larger efforts, including a Rust write-ahead log, a Bun migration and a replacement CI system.
Does Gill say people will stop reviewing software?
No. He expects less emphasis on reviewing every line of agent-written code and more human attention to plans, releases and final results. The interview also describes testing and staged rollouts as checks on agent work.
Are the reported results independently verified?
The source presents Gill’s account but does not include an independent audit of the spending, pull-request totals or CI cost and performance comparison. Those details should be treated as claims attributed to Gill.
Source: newsletter.pragmaticengineer.com
