PR review backlog pushed Harness to rebuild its Git repo

Harness has rebuilt its Code Repository to take agent traffic around the clock. It shipped the rebuild alongside AI Code Review, a product that decides which parts of a pull request still need a human reviewer. Field CTO Martin Reynolds described both in an interview with The New Stack published September 9, a few days after the launch.

Reynolds says he asks engineering leaders how they keep up with the pull requests their coding agents now produce. One recently answered with two words, “we’re not,” and explained that his team’s threshold for pushing code into production had dropped.

From 2x to 50x

Reynolds says he first hit the bottleneck during Harness’s early trials of GitHub Copilot and Amazon CodeWhisperer. “We were getting more PRs, but all the PRs were getting stuck,” he says, “and the test team was shouting, saying, we can’t keep up with all of this.”

That was a 1.5x to 2x increase in new code, and it pushed testing teams to the breaking point, he says. He now sometimes sees teams at 10x, and some claim 50x. “Imagine what that test team feels like right now.”

Among customers, Reynolds sees three groups. Some have raised their risk tolerance, some have a backlog they can’t manage, and most sit in between. “The somewhere in the middle, I think, is the most common,” he says. “We’re using some kind of another AI tool to help us in that space, but it doesn’t necessarily solve the problem.”

What the reviewer looks at first

Reynolds wants a reviewer who opens a pull request to see the most important changes first. He also wants the reviewer to be someone who has worked on that part of the codebase before, “not the person who did the prompt or wrote the code.” His example is a dependency update: “This other stuff is like 30 files because they updated a dependency. That’s less important in terms of getting eyes on.” Reviewers should “actually review what’s changing rather than a bunch of stuff that’s scaffolding around it.”

“It’s not just the model on its own,” he says. Harness spent “a good chunk of the last 12 months” building what it calls a software delivery knowledge graph: a map of a customer’s pipelines, deployments, incidents and policies. The reviewer queries it for context “at speed and not burn lots of tokens.” The company’s own example is a migration that got flagged because an earlier incident review had found that an unindexed CREATE INDEX statement locked a production table for 14 minutes.

Harness’s product page says the reviewer gates every pull request, cites delivery context such as incident history, and drafts PR descriptions from commits. By Harness’s own count, its engineers save more than 10,000 hours of manual review time a month. The reviewer also works on pull requests that already live on GitHub, so teams don’t have to move their code to use it. On the day Harness launched, GitHub’s Copilot code review started reviewing pull requests opened by bots, including those from GitHub’s own coding agent.

Agents don’t work nine to five

Harness customers on GitHub “would quite often genuinely send us screenshots of GitHub being down,” Reynolds says. He argues that 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, he says, “don’t work nine to five.”

GitHub’s outage on August 17 lasted nearly eight hours. In the postmortem, GitHub CTO Vlad Fedorov wrote that “a critical infrastructure component in our Central US data center failed to scale” when traffic hit a new peak. The postmortem puts GitHub at 2.9 billion commits a month, which works out to a little more than 1,000 a second on average. In April GitHub was already straining at 1.4 billion, so monthly volume roughly doubled in four months.

Harness has sold a repository service since 2023, when it launched Harness Code on top of its open-source Git project. TechTarget reported that the project, Gitness, joined the Harness platform in September 2023 as the successor to Drone CI, an open-source project Harness acquired in 2020. Development of Harness Open Source still happens on the main branch of the repository Drone came from, and Drone 2.x is maintained on its own branch.

Reynolds calls the rebuild “a ground-up AI-first repository that works for humans and AI.” He says it runs on Kubernetes across multiple clouds and regions and has been tested at thousands of commits per second. About 20 enterprise customers used it during the beta, and Harness has not named any of them. The thousands-per-second figure comes from load testing. GitHub’s figure of about 1,000 a second is an average of live production traffic, and Harness runs at nowhere near GitHub’s scale.

The product page adds the specs. Harness says the repository is scale-tested for more than 1,500 concurrent PRs, diffs touching more than 1,000 files, and terabytes of clone and fetch traffic a day. It promises no rate limiting, no bandwidth throttling and 99.99% uptime. The same branch, tag and push rules, CODEOWNERS files and secret scanning apply to human and agent contributors. RBAC (role-based access control) and OPA policy rules can limit what an agent identity is allowed to access, merge or deploy.

Agents connect through a Harness CLI or a Harness MCP server. The CLI’s structured commands return only the fields each operation needs, which keeps the token cost per call down. The MCP server is aimed at coding agents such as Cursor, Claude and Devin. Storage is 50 GB per account on the Free plan and 500 GB on paid plans, counting both Git and Git LFS. Harness offers one-click migration and mirroring, and pipelines can trigger on a push, a PR or a branch update.

What stays deterministic

Harness is also looking past the traditional delivery process. Reynolds says the capabilities for an autonomous delivery lifecycle already exist, then adds: “are organizations and companies ready for that? I’m not entirely sure.”

Either way, he says, deterministic tooling stays, and test results still come from the test runner. “There’s no need to rip those out and replace them. It’s like, where can you enhance them?”