Martin Reynolds, a field CTO at Harness, regularly poses a challenging question to engineering leaders: how are they managing the surge of pull requests generated by their AI coding agents? During a recent conversation, one leader responded with blunt candor: "we're not," and revealed that his organization had lowered its code production threshold as a result.

Reynolds spoke with The New Stack shortly after Harness unveiled a completely redesigned Code Repository and launched an AI Code Review offering. The timing coincided with GitHub's significant platform outage lasting nearly eight hours on August 17. The discussion centered on the review bottleneck that has emerged, the architectural decisions behind Harness's repository redesign for agent workloads, and which pipeline components must maintain deterministic behavior.

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

The Pull Request Avalanche

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

Reynolds first encountered this bottleneck when Harness began early experiments with 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 influx of new code—initially a 1.5x to 2x increase—pushed testing teams to their limits. Reynolds now observes teams experiencing 10x increases in pull requests, with some reporting 50x. At industry conferences, he found that drowning in pull requests had become a common refrain among engineering leaders.

When Reynolds speaks with customers, their responses fall into three categories: some have accepted higher risk levels, others face unmanageable backlogs, and most occupy the middle ground. "The somewhere in the middle, I think, is the most common," Reynolds says. "We're using some kind of another AI tool to help us in that space, but it doesn't necessarily solve the problem."

Focusing Reviews on Meaningful Changes

Reynolds advocates for reviewers to prioritize substantive modifications over boilerplate alterations. When opening a pull request, reviewers should encounter the most critical changes first, and ideally, those reviewers should be individuals with prior experience in that codebase section—not necessarily the person who submitted the prompt or authored the code.

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

Beyond the AI model itself, Harness invested heavily over the past year in constructing 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 prior incident investigation had identified an unindexed CREATE INDEX statement that had locked a production table for 14 minutes.

According to Harness's own metrics, its engineers have recovered more than 10,000 hours of manual review time monthly. On the day Harness launched its new offering, GitHub's Copilot code review feature began examining pull requests submitted by bots, including its own coding agent.

Building for Round-the-Clock Agent Operations

Harness customers using GitHub have frequently sent Reynolds screenshots documenting GitHub outages. Reynolds attributes GitHub's infrastructure challenges to its 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." Agents, by contrast, "don't work nine to five."

Harness has offered a repository service since 2023, when it introduced Harness Code built on its open-source Git foundation. Reynolds describes 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, has been stress-tested at thousands of commits per second, and is currently used by approximately 20 enterprise customers in beta, though Harness has not publicly identified them.

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

While Harness does not operate at GitHub's scale, this distinction may represent an advantage for its enterprise customers. Reynolds suggests that the technical capabilities for an autonomous delivery lifecycle already exist today, yet questions whether "organizations and companies ready for that? I'm not entirely sure."

Reynolds emphasizes that deterministic tooling must remain intact, with test results 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 most immediately accessible entry point for most teams. It operates on pull requests already hosted on GitHub, whereas migrating a repository constitutes a lengthy undertaking for most enterprises.

The engineering leader who told Reynolds "we're not" does not necessarily require a new Git hosting platform to shift that answer. What he needs is tooling that distinguishes which files in a pull request demand human attention and which 30 arrived as part of a dependency update. This represents a narrower objective than establishing a fully autonomous delivery lifecycle, but for the present moment, it may prove more immediately valuable.

Source: The New Stack