GitHub Is Rebuilding Git Infrastructure for a World Where AI Writes the Code

A human developer might push code a dozen times on a productive day. An AI coding agent can push hundreds of times in a single hour, across multiple branches simultaneously, without stopping to review or deliberate. That gap is real. It is why GitHub is rebuilding its infrastructure from the ground up.

GitHub’s engineering team published a detailed technical writeup this week on their effort to build what they call agent-scale Git infrastructure. The core problem is not storage. It is contention. Every time a write happens in Git, the system places a lock on the repository, a safeguard that prevents two operations from colliding and corrupting the codebase. That locking model was designed for human teams pushing code a handful of times per day. When a hundred AI agents are pushing simultaneously, the entire repository queues up behind each write one at a time, and throughput collapses.

The compounding factor is cross-region latency. GitHub has been migrating to Azure, and 99.99999% of repositories have already made the move. The remaining repositories are the highest-traffic ones on the platform, and they could not complete the migration because the 25-millisecond round trip to Azure’s data centers made the locking problem worse. A slow replica in a distant region sets the pace for every write. The busiest repos hit their ceiling fast.

The new design uses per-reference quorums to solve this. A quorum is a majority-agreement system where an operation completes once enough replicas confirm it, rather than waiting for every single one to respond. Per-reference means that each individual Git reference, such as a specific branch, gets its own lock rather than locking the entire repository. Two agents pushing to different branches at the same time no longer block each other at all. Internal tests showed three to four times higher write throughput for repositories with concurrent activity. That is a material difference, not a marginal one.

The read layer is being decoupled from writes entirely. In the new architecture, read capacity can scale through object storage independently of the write infrastructure, meaning you can absorb spikes in read traffic without reconfiguring or expanding the write tier. For large organizations managing shared codebases across many teams, that kind of independence is operationally significant.

The reason all of this is happening now is the agentic shift. GitHub’s own traffic data shows exponential growth driven by AI agents rather than human developers. Code review agents, automated testing tools, and AI coding assistants all interact with Git constantly, at a frequency no human workflow was built to sustain.

If your team is already using AI-assisted development tools, or planning to, your internal Git workflows will face the same pressure at a smaller scale. The bottlenecks shift when agents join the development process. Your repository structure, branching strategy, and CI/CD pipelines may need to be revisited. Knowing that GitHub is solving this problem at the platform level helps you understand what to watch for in your own infrastructure.

Want to explore how AI-assisted development could benefit your engineering team? Let’s talk.