AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

CodeRabbit founder and CEO Harjot Gill says he shipped more than 5,000 pull requests in one month using AI agents, while the company spends millions of dollars a month on AI. In a podcast interview, he described building internal software systems with agents and argued that teams need layered testing and human review focused more on plans and releases than individual pull requests.

Harjot Gill, founder and CEO of AI code-review company CodeRabbit, said his team shipped more than 5,000 pull requests in one month with help from AI agents, describing a software-building approach that shifts some engineering work from writing code to directing and checking automated systems. In an interview on The Pragmatic Engineer Podcast, Gill also discussed the company’s high AI spending and argued that reliable agent-built software needs multiple layers of verification.

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

Gill said the pull requests included substantial maintenance and infrastructure work, such as fixing flaky tests, improving continuous integration and refining caching. He described starting agents during his commute or in the evening, then reviewing their output. In some cases, he directed agents to improve the user interface. The source does not give a breakdown of how many pull requests were routine maintenance, product work or other changes.

Examples he cited include a write-ahead log built in Rust that ran in production after an agent worked on it for six weeks, a weekend migration of CodeRabbit services to Bun, and a replacement CI system. Gill said the company chose to build its own CI system after receiving a quote of $13 million per year for a service that could handle its pull-request volume. According to the source, the agent-assisted replacement took about a week and cost 50% less than the quoted price while delivering similar performance. Those cost and performance figures are Gill’s account.

The source describes CodeRabbit as having 40 engineers and spending millions of dollars per month on AI. Gill put his own monthly AI spend at roughly $250,000 to $275,000 and estimated a cost of about $150 per pull request. The episode’s host contrasted that scale with spending of roughly $200 to $2,000 per developer per month that the host said they hear about elsewhere. These figures come from the interview and are not presented as an independently audited comparison.

At a glance
reportWhen: Discussed in a recent episode of The Pr…
The developmentIn a Pragmatic Engineer podcast interview, CodeRabbit CEO Harjot Gill detailed how his team uses AI agents to produce thousands of pull requests and how it verifies their work.
AI Engineering · Briefing

How Harjot Gill Builds Software Factories

CodeRabbit’s CEO describes using AI agents to produce software at high volume, with people shifting their attention toward plans, verification and releases.

5,000+
pull requests shipped in one month, Gill said
40
engineers at CodeRabbit
$250K–$275K
Gill’s estimated monthly AI spend

Agents expand the work

Gill said agents handled infrastructure and product tasks, while he reviewed their output.

Infrastructure

Fix and improve systems

Examples included flaky test fixes, CI improvements and caching work.

Product

Build through direction

Gill sometimes directed agents to improve the user interface.

Verify at every stage

Gill’s approach combines automated checks with human oversight throughout development and release.

1
Model and simulate

Use formal methods such as TLA+ and Lean, plus local simulations.

2
Test in development

Check changes with emulated APIs and development deployments.

3
Roll out gradually

Monitor logs and performance during staged production releases.

Review moves toward releases

When agents create many small changes, Gill says teams can review combined work and releases rather than centering every pull request.

Agents handle
  • Producing and refining changes
  • Running checks and simulations
  • Supporting maintenance work
People oversee
  • Plans and combined changes
  • Release readiness and rollout
  • Accountability for outcomes
“Human code review is dead, long live the human review!”— Harjot Gill
CaveatThe pull request, spending and cost claims come from the interview and have not been independently verified.
The takeaway
Agent-heavy engineering depends on layered verification and accountable human review across the release cycle.

From Pull Requests to Release Review

Gill’s account offers a concrete example of how AI agents can change the unit of engineering work. If agents produce many small changes, a team may spend less time reviewing each individual pull request and more time checking the overall plan, combined changes and resulting release. Gill said CodeRabbit engineers review releases made up of multiple pull requests rather than treating every agent-generated pull request as the main review point.

That approach does not remove the need for human oversight. It changes where oversight happens and makes testing, deployment controls and accountability more important. A larger volume of changes can speed up maintenance and product development, but it can also make errors harder to spot if teams rely on code review alone. Gill’s examples show why his view of a “software factory” includes systems that test, simulate and monitor changes—not just agents that generate code.

The spending figures also illustrate why costs may vary widely between companies. Gill and the podcast host framed CodeRabbit’s AI spend as a possible form of research and development: the company builds developer tools and may use its own agents to develop products it can sell. That reasoning may not apply to firms with established products and customers who will not pay more for additional features. The interview presents a company case study, not evidence that comparable spending will pay off for every startup.

Products Worth Considering

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, White Smoke

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, White Smoke

Kindle Paperwhite Signature Edition 32GB (2024 Model) - Metallic Black

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

As an affiliate, we earn on qualifying purchases.

CodeRabbit’s Shift Toward Agents

Gill told the podcast that he expected in 2021 that AI would write most code. While working at FluxNinja, he had seen code review become a bottleneck as tools such as GitHub Copilot began speeding up code completion. He said he underestimated the market for AI code review, believing at the time that demand would mainly come from larger companies that already bought code-verification products. He now sees the need as much broader.

The discussion links that early experience to CodeRabbit’s current use of agents. The source says Gill has spent roughly 30 years building software and infrastructure startups. It does not provide independent performance data for the company’s workflow, nor does it establish that the same process is suitable for teams with different products, risks or engineering resources.

Gill’s view is that agents need to be able to improve through feedback—what he described as “hill climbing”—and verify their own work. The verification methods cited include formal models such as TLA+ and Lean, local simulations that test different interleavings of concurrent operations, emulated APIs, development deployments and staged production rollouts. He said a TLA+ model identified race conditions that code review had missed. He also described agents checking logs and performance during progressive rollouts before a wider release proceeds.

“Human code review is dead, long live the human review!”

— Harjot Gill, describing his view of engineering review

Limits of the Spending and Output Figures

The account is based on a podcast interview and does not include independent verification of the pull-request count, AI spending, CI cost comparison or performance claims. The source also does not say how the 5,000-plus pull requests were counted, how many were merged, or what portion were generated mostly by agents versus substantially directed or edited by engineers.

It is also unclear how CodeRabbit measures the quality and long-term maintenance cost of this volume of changes, or how its internal review exemptions work. The supplied source ends mid-sentence while describing a “review exempt list” of developers, so it does not establish who qualifies, what permissions the list grants or what safeguards apply. No episode date is given, which limits precision about when the figures were recorded.

Verification Across the Release Cycle

Gill’s account points to verification at several stages as the next practical test for agent-heavy engineering: formal analysis where appropriate, repeatable simulations, development environments, and feature-flagged releases monitored before wider deployment. The interview does not announce a new product launch or set out a future rollout timetable.

For readers evaluating this approach, the open questions are whether the company’s reported speed and savings persist over time and how teams balance agent autonomy with human responsibility. The source offers Gill’s current methods and reasoning, but does not provide subsequent results or independent comparisons with conventional engineering workflows.

Key Questions

What did Harjot Gill say CodeRabbit shipped in one month?

Gill said he shipped more than 5,000 pull requests in a month with agents. The source does not give a count of how many were merged or describe the counting method.

How much does CodeRabbit spend on AI?

The source says CodeRabbit spends millions of dollars per month on AI. Gill estimated his own monthly spend at about $250,000 to $275,000; the figures are from the interview, not an independent audit.

Does Gill think human review is no longer needed?

No. Gill argues that engineers may review plans and releases containing multiple pull requests rather than reviewing each agent-generated pull request individually. His approach still includes human oversight.

How does CodeRabbit check work produced by agents?

Methods described in the interview include formal models, concurrency simulations, API emulation, development deployments and staged production rollouts. Gill said a TLA+ model found race conditions that code review had missed.

Are CodeRabbit’s cost and performance results independently confirmed?

The supplied source attributes the reported costs, savings and performance to Gill. It provides no independent verification or comparative study of the company’s results.

Source: newsletter.pragmaticengineer.com

You May Also Like

The New Push for Energy-Efficient Software Development

AIThis post was created with the assistance of artificial intelligence (AI).The new…

Learning to Code in 2025: AI Tutors and New Pathways

Jump into the future of coding education in 2025, where AI tutors and innovative pathways unlock your potential—discover how these tools will transform your learning journey.

Why Passkeys Are Finally Moving Into Mainstream Apps

Unlock the future of secure, password-free login with passkeys—discover why they’re finally becoming mainstream and how they can protect your accounts better.

How Government Privacy Rules Are Changing Product Analytics

Obliged by new government privacy rules, discover how to adapt product analytics to stay compliant and protect user data effectively.