Harness has replaced its original Git service with a ground‑up, AI‑first Git repository that runs on Kubernetes across multiple clouds and regions, and it ships an AI‑driven code‑review assistant. The change targets the surge of pull requests generated by coding agents, which traditional Git hosts struggle to handle at scale.
Why the Change Was Needed
Engineering teams that adopted Copilot or CodeWhisperer reported a 1.5‑2× increase in new code, with some customers seeing 10× or even 50× growth in pull‑request volume. Test teams could not keep up, and the existing Git platform showed latency during a recent multi‑hour outage at a major provider. Harness observed that the conventional Git workflow—designed for small human teams with short‑lived PRs—does not accommodate agents that operate continuously.
AI‑First Git Repository Architecture
The rebuilt service is built on the open‑source Git project but is re‑hosted in a Kubernetes‑based control plane that spans several cloud providers and geographic regions. It is engineered to sustain “thousands of commits per second” in testing, and it integrates a software‑delivery knowledge graph that links pipelines, deployments, incidents, and policy data. This graph supplies contextual information to the AI code‑review engine without consuming excessive model tokens. The repository still accepts pull requests that live on external Git hosts, allowing incremental migration rather than a wholesale move.
Operational Implications
Running a custom Git service shifts operational responsibility to the consumer. Practitioners must provision and maintain a multi‑cloud Kubernetes cluster, monitor commit‑rate metrics, and ensure that the knowledge‑graph data pipeline stays in sync with existing CI/CD tools. Deterministic steps such as test execution remain external to the repository; the AI review layer augments rather than replaces them. The reported savings of more than 10,000 hours of manual review per month illustrate the potential efficiency gain, but they also imply a reliance on the AI assistant’s accuracy and on the underlying knowledge graph’s freshness.
Related CloudNinjas coverage: DevOps.
What This Means For Practitioners
Teams should evaluate whether their pull‑request volume justifies a dedicated, AI‑first Git backend. Key actions include:
- Benchmark current commit and PR rates against the “thousands per second” target to gauge scaling needs.
- Assess the effort required to operate a multi‑cloud Kubernetes deployment, including security hardening and observability.
- Map existing deterministic tooling (test runners, policy checks) to the new repository to confirm they remain compatible.
- Plan a phased migration that keeps critical repositories on the legacy host while pilot‑testing the AI review workflow.
Monitoring the AI assistant’s recommendations for false positives and ensuring the knowledge graph stays up‑to‑date will be essential to maintain code quality and avoid hidden risk. The shift also opens a window to experiment with more autonomous delivery pipelines, but organizations should retain manual verification points until confidence in the AI layer is proven.