Greptile puts numbers on the AI pull request spam problem

Why it matters

AI coding agents make pull requests cheaper to create, which shifts value toward review, execution, reputation and maintainer trust. Greptile is trying to own that control point.

Daksh Gupta (@dakshgup) built Greptile around a bet that code review, not code generation, would become the next bottleneck in software. A May 8 analysis by Greptile engineer Rahul Bathija gives that bet a concrete case study: OpenClaw, a fast-growing open source repository that Greptile reviews, saw pull requests rise from about two per week in December to 3,400 per week by February, according to the company.

The more important number is what happened after the volume arrived. Bathija wrote that OpenClaw's merge rate fell from roughly 48% before the spike to below 9.3% after it. One contributor submitted 106 pull requests in a single day, with a median of three seconds between submissions. Greptile's line is blunt: "PR spam today looks like email spam in the early 2000s."

That is not a neutral academic finding. Greptile sells the thing this problem points toward: an AI code reviewer with repo context, now paired with TREX, its execution layer for running code during review. In its TREX launch post, Greptile says this moves the product beyond static diff-reading toward executable evidence, including logs, screenshots, API traces and scripts. The OpenClaw data is therefore both a warning to maintainers and a sales argument for Gupta's broader thesis: once agents make code cheap to produce, the scarce resource is trustworthy validation.

From codebase search to validation layer

Greptile did not start as a pull request spam company. According to Greptile's seed announcement, the team set out to see whether large language models could answer hard questions about real codebases the way a senior engineer would. Greptile announced a $4.1 million seed round in June 2024 led by Initialized Capital.

The founder backdrop matters because the product has followed the team's own learning curve. Greptile's early pitch was codebase understanding: index a large repo, let developers ask questions in plain English, and expose that capability through an API. TechCrunch reported in 2024 that the product had grown to 500 paying customers after coming through Y Combinator. Greptile's YC profile lists Gupta, Soohoon Choi and Vaishant Kameswaran as founders.

Greptile has since narrowed the wedge. Greptile now describes itself as AI agents that review pull requests with full codebase context. In September 2025, Greptile said it had raised a $25 million Series A led by Benchmark.

The pricing tells the same story in operational terms. Greptile's billing documentation prices code review at $30 per active developer per month, including 50 reviews, with additional reviews billed at $1 each. If AI coding agents push more changes into the review queue, Greptile's commercial opportunity grows with the pressure on maintainers.

OpenClaw is the lab, not the whole market

Bathija's OpenClaw post is valuable because it separates volume from usefulness. In Greptile's dataset, contributors with more history performed better: first-time contributors had an 8.2% merge rate, contributors with two to five pull requests had a 10.3% merge rate, and contributors with more than five had an 18.6% merge rate. That is a small but telling signal that maintainers are already applying sender reputation, even if the tooling has not caught up.

The comparison to email spam is useful because the failure mode is economic. When the cost of sending a message collapses, inboxes need identity, reputation and filtering. When the cost of producing a patch collapses, maintainers need the same primitives for pull requests. The code may compile. The diff may look plausible. The burden shifts to proving that the sender understood the project, the context and the consequences.

OpenClaw also showed a second problem: agent-generated sameness. Greptile says four contributors submitted pull requests with the exact same SearXNG feature title, six people independently fixed the same Brave Search locale bug, and five people independently found the same timeout deadlock in the agent runner. The old open source ideal was that more contributors meant more perspective. The new risk is that more contributors using the same models and prompts can produce the same solution shape at high speed.

That matters for more than etiquette. Bathija's post says feature pull requests in the OpenClaw data merged at 9%, while refactors merged at 35%. The interpretation Greptile wants readers to draw is that the durable work is not typing code but understanding the system well enough to make the right change. That is founder-friendly for the open source community, not defeatist: agents can expand who participates, but maintainers still need ways to distinguish informed contribution from automated noise.

Reputation is already becoming infrastructure

The maintainers are not waiting for vendors to finish the stack. Bathija points to Vouch, Mitchell Hashimoto's trust-management project for open source contributors. The GitHub repo describes Vouch as a community trust management system based on explicit vouches. Greptile frames it as the open source equivalent of a sender reputation score.

Hashimoto's own arc is a warning sign for platforms. He later wrote that Ghostty is leaving GitHub, after years of deep personal investment in the platform. Bathija notes that Vouch had been working well for Ghostty before that decision. Whether Vouch itself becomes broadly adopted is less important than the direction of travel: contribution workflows are moving from open-by-default to trust-mediated.

That shift creates room for products like Greptile, but it also creates tension. Greptile's product can review a diff, run code through TREX, learn team standards from comments and flag issues. It cannot fully solve contributor identity, project governance or maintainer burnout. Those are social systems with software interfaces. The winners in this market will be the companies that understand that distinction instead of pretending every review problem is a model-quality problem.

The competitive pressure is validation

Greptile's latest argument lands in a category that is no longer empty. GitHub has the native advantage because pull requests already live there. Cursor, CodeRabbit, Qodo, Graphite and others are pushing adjacent workflows around code review, agent feedback, testing and developer productivity. Greptile's answer is independence: Gupta has repeatedly framed the company as an auditor rather than another code generator. In an August 2025 post, he wrote that Greptile's role is to catch bugs in pull requests for software teams, and that its positioning is deliberately not to generate the code it reviews.

That is the cleanest version of the business. If every engineer and every agent can produce code faster, the reviewer becomes the control point. Reviewers decide what ships, what gets blocked, what gets routed back to the author and what becomes training signal for the next agent loop. Greptile is trying to own that control point before it gets absorbed by the platforms where developers already work.

OpenClaw is not proof that all open source is drowning in AI slop. It is one repository, filtered through Greptile's own visibility, with methodology the company has not exposed as a reusable public dataset. But it is a useful early stress test because the numbers are extreme enough to show what breaks first: maintainer attention, duplicate work, and confidence that a pull request represents a thoughtful contributor rather than a cheap send.

For Gupta and Greptile, that is the market opening. Greptile began by helping developers understand large codebases. The agent era is forcing a sharper question: who, or what, gets trusted to say a change is safe enough to merge?