PR review backlog pushed Harness to rebuild its Git repo

Harness has rebuilt its Code Repository and launched an AI Code Review product that ranks which files in a pull request still need human attention, work the company says saves its own engineers more than 10,000 hours of manual review time a month. Field CTO Martin Reynolds described the review bottleneck behind both products in an interview with The New Stack.

At a glance

• Reynolds says early coding-assistant trials brought a 1.5x to 2x rise in new code that broke test teams, and he now sometimes sees teams at 10x their old volume, with some claiming 50x.
• The rebuilt repository runs on Kubernetes across multiple clouds and regions, was tested at thousands of commits per second, and went through beta with about 20 enterprise customers that Harness has not named.
• The AI Code Review works on pull requests already hosted on GitHub, so teams can rank which files still need human attention without first migrating a repository, a long project at most enterprises.

The bottleneck has moved. Generating code stopped being the constraint once agents took it over, and review capacity became the limit, which is why a repository rebuild and a reviewer that ranks diffs read as answers to the same problem. Harness appears to be betting that enterprises adopt the reviewer first, since it requires no migration, and that the repository follows only where Git hosting itself has turned into an operational risk.

Reynolds says an early 1.5x to 2x jump in code volume now reaches 10x at some teams

Reynolds first hit the bottleneck during Harness's early trials of GitHub Copilot and Amazon CodeWhisperer, when the team was getting more pull requests but all of them were getting stuck and the test team said it could not keep up. That 1.5x to 2x increase in new code pushed testing teams to the breaking point; he now sometimes sees teams at 10x their old volume, with some claiming 50x.

He describes customers splitting three ways: some have raised their risk tolerance, some carry a backlog they cannot manage, and most sit in the middle, using another AI tool that does not necessarily solve the problem. One engineering leader, asked how his team keeps up, answered "we're not," and said his threshold for pushing code into production had dropped.

A knowledge graph flagged a migration tied to a 14-minute production table lock

It is not just the model on its own, Reynolds says. Harness spent a good chunk of the last 12 months building a software delivery knowledge graph, a map of a customer's pipelines, deployments, incidents and policies, so the reviewer can pull context at speed without burning lots of tokens.

The company's own example is a migration flagged because an earlier incident review had found an unindexed CREATE INDEX statement locking a production table for 14 minutes. Reynolds also argues that 30 files arriving with a dependency bump matter less than the code that actually changed, and that reviewers should come from whoever has worked on that part of the codebase.

GitHub handles 2.9 billion commits a month, a little more than 1,000 a second

Harness customers on GitHub sent screenshots of GitHub being down, Reynolds says. He attributes the strain to a platform built for people, teams of maybe 10 or 15 changing code and opening pull requests that live for a few hours to a couple of days, while agents do not work nine to five.

GitHub CTO Vlad Fedorov's postmortem on the nearly eight-hour August 17 outage said a critical infrastructure component in the Central US data center failed to scale as traffic hit a new peak. GitHub now handles 2.9 billion commits a month, a little more than 1,000 a second on average.

Harness has sold a repository service since 2023, when it launched Harness Code on top of its open-source Git project. Reynolds calls the rebuilt version a ground-up AI-first repository for humans and AI. The day Harness launched, GitHub's Copilot code review began reviewing pull requests opened by bots, including its own coding agent.

Whether autonomy comes next

Reynolds says the capabilities for an autonomous delivery lifecycle exist today, but he is not sure organizations are ready for one, and he sets no date for that shift. He argues deterministic tooling should survive it, with test results still coming from the test runner rather than a model, and asks instead where existing tools can be enhanced.