Harness overhauls Git to handle relentless AI agent code submissions

Engineering teams are facing an unexpected crisis: their AI coding agents are generating so much code that human reviewers can't possibly keep pace. One engineering leader recently summed up the predicament in two words: "we're not."

This glut of pull requests prompted Harness to rearchitect its Git repository from the ground up, introducing both a rebuilt Code Repository and an AI Code Review product designed to handle what traditional platforms never anticipated—machines working around the clock.

The review bottleneck nobody expected#

When Martin Reynolds, a field CTO at Harness, began working with early deployments of GitHub Copilot and Amazon CodeWhisperer, the pattern became clear: more code generated meant more work backed up. Testing teams struggled to validate an additional 1.5 to 2x volume of submissions, but in some cases, teams now report 10x or even 50x increases.

Conversations with engineering leaders reveal three different responses to the onslaught. Some teams have simply accepted higher risk by lowering quality thresholds. Others face unmanageable backlogs. Most fall somewhere in between—using AI tools to help manage the flood but finding they don't solve the fundamental problem of human reviewers drowning in submissions.

Making reviews smarter, not just faster#

The real challenge isn't reviewing code faster—it's reviewing the code that matters. Reynolds points out that when reviewers open a bloated pull request, they're often seeing scaffolding and boilerplate changes (like dependency updates across 30 files) mixed in with the actual logic changes.

Effective code review, he argues, should prioritize substantive modifications and route reviews to developers who previously worked in those specific code areas—not whoever wrote the prompt. Harness spent the past year building what it calls a software delivery knowledge graph, mapping customer pipelines, deployments, incidents, and policies to give reviewers contextual information without burning excessive API tokens.

The impact has been measurable: Harness reports its engineers saved over 10,000 hours monthly on manual review work alone.

Why traditional Git wasn't built for this#

GitHub and other platforms were designed for human teams—typically 10 to 15 developers creating pull requests that sit for hours or days before merging. Agents operate on a completely different schedule. They don't work nine-to-five; they generate code continuously.