Engineering teams using AI coding assistants are now facing a 1.5x to 2x surge in pull requests, with some organizations reporting increases as high as 10x to 50x, according to a report published by The New Stack following the launch of Harness's rebuilt Code Repository and new AI Code Review product. The report, based on conversations with Martin Reynolds, a field CTO at Harness, describes how the explosion of machine-generated code is overwhelming review and testing teams. Reynolds told The New Stack that one engineering leader recently summed up his team's ability to keep pace with the flood of pull requests in two words: "we're not."
The bottleneck first emerged during Harness's early trials of GitHub Copilot and Amazon CodeWhisperer, Reynolds explained, when increased pull requests began piling up and testing teams couldn't maintain their review pace. What began as a 1.5x to 2x rise in new code submissions pushed testing teams to their breaking point, and the problem has since intensified dramatically. In hallway conversations with engineering leaders at a recent conference, Reynolds found that drowning in pull requests had become a recurring theme. The response among customers typically falls into three camps: some have raised their risk tolerance and lowered the bar for pushing code into production, some are struggling with unmanageable backlogs, and most occupy a middle ground where they're using additional AI tools to assist with reviews but haven't truly solved the underlying problem.
According to Reynolds, the ideal review process should surface the most critical changes first, with reviewers drawn from whoever has previously worked on that section of the codebase rather than whoever wrote the prompt or generated the code. He noted that reviewers should "actually review what's changing rather than a bunch of stuff that's scaffolding around it," pointing to scenarios where 30 files might change simply because a dependency was updated. The report states that Harness built what it calls a software delivery knowledge graph over the past 12 months, mapping customer pipelines, deployments, incidents, and policies to provide context quickly without consuming excessive tokens. By the company's own count, its engineers saved more than 10,000 hours of manual review time monthly.
The core issue, Reynolds argues, is that GitHub and similar platforms were originally designed for human teams of perhaps 10 to 15 people creating pull requests that would remain open for a few hours to a couple of days. But coding agents "don't work nine to five," and the continuous flow of machine-generated code is straining infrastructure that wasn't architected for that pattern. Harness rebuilt its repository service as what Reynolds describes as an AI-first system running on Kubernetes across multiple clouds, tested at thousands of commits per second. The report notes that GitHub's August 17 outage lasted nearly eight hours, with GitHub CTO Vlad Fedorov's postmortem attributing it to a critical infrastructure component that failed to scale as traffic hit a new peak. While GitHub now handles 2.9 billion commits monthly—roughly 1,000 per second on average—Reynolds suggests that for enterprise users, Harness's smaller scale may actually be an advantage. The review tool launched by Harness works on pull requests already hosted on GitHub, meaning teams don't need to migrate their entire repository to change their answer from "we're not" keeping up. For the engineering leader Reynolds spoke with, the immediate need isn't a new Git host but something that tells reviewers which files in a pull request need human attention and which 30 arrived with a routine dependency update. That's a narrower promise than an autonomous delivery lifecycle, but it's likely the more immediately useful one for teams currently underwater. The pressure to ship faster while maintaining quality will only intensify as coding agents become standard tools rather than experimental add-ons. Organizations that can't solve the review bottleneck may find themselves choosing between unacceptable risk and unsustainable workload—neither of which leads to competitive advantage in the long run.