Harness Overhauls Git Infrastructure to Handle AI Agent Pull Request Surge

Martin Reynolds, a field CTO at Harness, regularly poses a challenging question to engineering leaders: how are they managing the volume of pull requests generated by their AI coding agents? When one leader responded with just two words—"we're not"—he revealed that his organization had lowered its standards for pushing code to production, according to Reynolds's conversation with The New Stack.

The discussion took place shortly after Harness unveiled a completely redesigned Code Repository and launched an AI Code Review product. The timing coincided with GitHub's nearly eight-hour service disruption on August 17. Reynolds and The New Stack explored how the review bottleneck developed, what prompted Harness to reconstruct its Git repository for agent-driven workloads, and which pipeline components should maintain deterministic behavior.

https://www.youtube.com/embed/Cm0rOsy-y0U?si=PyA3_ZGqZuj0dEYk

Drowning in pull requests

https://player.simplecast.com/6ced3fc9-3d14-4f04-8990-319411c0950a?dark=true

Reynolds first encountered this bottleneck when Harness began testing GitHub Copilot and Amazon CodeWhisperer. "We were getting more PRs, but all the PRs were getting stuck," he recalls, "and the test team was shouting, saying, we can't keep up with all of this."

Imagine what that test team feels like right now.

Martin Reynolds, Harness Field CTO

The initial surge brought a 1.5x to 2x increase in new code submissions, straining testing operations beyond capacity. Reynolds now observes teams experiencing 10x increases, with some reporting 50x. During conversations with engineering leaders at industry conferences, the challenge of managing pull request volume emerged as a consistent concern. His customer interactions reveal three distinct patterns: some organizations have accepted greater risk, others face unmanageable backlogs, and the remainder occupy a middle ground.

"The somewhere in the middle, I think, is the most common," Reynolds observes. "We're using some kind of another AI tool to help us in that space, but it doesn't necessarily solve the problem."

Review what's changing, not the scaffolding

When reviewers examine a pull request, they should encounter the most significant modifications first, Reynolds argues. Ideally, reviewers should be individuals with prior experience in that section of the codebase, rather than whoever authored the code or submitted the prompt.

"This other stuff is like 30 files because they updated a dependency. That's less important in terms of getting eyes on," he explains. Reviewers should "actually review what's changing rather than a bunch of stuff that's scaffolding around it."

The solution extends beyond the AI model itself. Over the past 12 months, Harness developed what it terms a software delivery knowledge graph—a comprehensive representation of customer pipelines, deployments, incidents, and policies. This enables reviewers to access context "at speed and not burn lots of tokens." Harness cites one example: a migration that was flagged because a previous incident investigation had identified an unindexed CREATE INDEX statement that had frozen a production table for 14 minutes.

According to Harness's internal metrics, its engineers recovered more than 10,000 hours monthly in manual review work. On the day Harness launched its product, GitHub's Copilot code review feature began analyzing pull requests created by bots, including GitHub's own coding agent.

Agents don't work nine to five

Harness customers using GitHub have frequently sent Reynolds screenshots documenting GitHub outages. He attributes GitHub's infrastructure challenges to the platform's original design philosophy: "GitHub was ultimately built for people, teams of maybe up to 10, 15, who are changing code, creating pull requests. Those pull requests will be there for a few hours to maybe a couple of days." AI agents, by contrast, "don't work nine to five."

Harness has operated a repository service since 2023, when it introduced Harness Code atop its open-source Git foundation. Reynolds characterizes the rebuilt version as "a ground-up AI-first repository that works for humans and AI." The infrastructure runs on Kubernetes across multiple clouds and regions, with testing demonstrating capacity for thousands of commits per second. Approximately 20 enterprise customers participated in the beta phase, though Harness has not disclosed their identities.

GitHub CTO Vlad Fedorov's analysis of the August 17 outage attributed the failure to "a critical infrastructure component in our Central US data center failed to scale" when traffic reached unprecedented levels. GitHub currently processes 2.9 billion commits monthly, averaging slightly above 1,000 commits per second.

Harness does not operate at GitHub's scale, but for its enterprise users, that distinction may represent a competitive advantage. Reynolds contends that the capabilities for an autonomous delivery lifecycle already exist, yet questions whether "organizations and companies ready for that? I'm not entirely sure."

Reynolds emphasizes that deterministic tooling must remain in place, with test outcomes continuing to originate from the test runner. "There's no need to rip those out and replace them. It's like, where can you enhance them?" The code review component represents the feature most teams will adopt initially. It operates on pull requests already stored on GitHub, whereas migrating repositories constitutes a lengthy undertaking for most enterprises.

The engineering leader who told Reynolds "we're not" does not require a new Git host to shift his answer. What he needs is a tool that directs his reviewers toward files requiring human attention while identifying the 30 files modified solely due to a dependency update. This represents a more modest objective than a fully autonomous delivery lifecycle, yet for the present moment, it addresses the more pressing need.

Source: The New Stack