MONDAY, NOVEMBER 16, 2026 Tech news, minus the noise.
N News Cast
LATEST
Home  /  Dev
Dev

GitHub Rebuilt Git Storage for 5,000 AI Agents Pushing at Once

GitHub unveiled Project Rubicon at GitHub Universe, a new Git storage architecture that handled 5,000 concurrent agents and 7,200 pushes per second in internal tests. It is built for a world where agents write most of the code.

GitHub Rebuilt Git Storage for 5,000 AI Agents Pushing at Once
Source: via Wikimedia Commons (CC BY-SA 4.0)

Git was built for humans. Agents push a thousand times faster

GitHub unveiled a new Git storage architecture at GitHub Universe in early November 2026, and the numbers explain why it was needed. In internal testing, the system supported 5,000 concurrent agents working on a single repository, sending 7,200 pushes per second, while serving read traffic with consistently low latency.

The architecture is codenamed Project Rubicon, and GitHub says it is the foundation for agentic development at the scale of hundreds of millions of developers. The busiest repositories today receive at most dozens of writes per second, and only in short bursts. GitHub expects those bursts to become a sustained rate within a year, as AI agents work in parallel around the clock.

That is a polite way of saying the old system was about to break. Git was designed in 2005 for humans committing code a few times an hour. A fleet of coding agents can generate that much activity before lunch. GitHub Rubicon storage is the rebuild that lets the platform survive its own success.

Why the old design hits a wall

The problem is coordination. In a traditional Git hosting setup, writes to a repository need to agree on what the current state is before they can land. That works fine when writes are rare. It becomes a bottleneck when thousands of agents push at once, because every push has to wait its turn to be ordered.

Reads make it worse. Every push fans out to CI pipelines, code review tools, and other agents that read the repository. If writes slow down reads, the whole system feels sluggish to the humans still using it. GitHub's design goal was explicit: accept thousands of writes per second, and let reads scale on demand without getting dragged down.

The company started designing Rubicon in May 2026, after a bridge step that moved 99.99999% of repositories into Azure, scoped the write quorum to individual references, and stood up virtual replicas backed by Azure Blob Storage. Rubicon is the architecture that builds on that foundation.

How Rubicon actually works

Rubicon rests on two principles: minimize coordination, and decouple storage from compute. The key insight is that Git objects are immutable and content-addressed, which means writers can store them without agreeing with anyone first. Only reference updates need agreement, and those are a tiny fraction of the work.

In practice, Git objects live as packs in Azure Blob Storage. References and the pack manifest live in a separate replicated log service called Ref-CAS, where a small consensus group orders reference updates independently. Writes are handled by virtual replicas, disposable compute nodes that index incoming packs, upload them to Blob Storage, and then commit a reference transaction to Ref-CAS.

Compaction, the housekeeping that keeps the pack count from exploding, scales horizontally across shards. GitHub says the architecture showed a flat latency curve as writes climbed, which is the result that matters most: performance that does not degrade as load grows.

The road so far, and what is still preview

Rubicon is available in a limited private preview now, with a broader rollout planned for early next year. GitHub has not announced a date for full rollout, and the architecture has not replaced the current Git serving stack yet.

That matters for how to read the headline numbers. Five thousand agents and 7,200 pushes per second are capacity ceilings under synthetic stress, not observed production traffic. The real test comes when sustained agent load hits real repositories with real humans trying to merge at the same time.

Still, the direction is clear. VS Code 1.146 gave AI coding agents an inbox, and .NET 11 shipped with an MCP server template. The tooling is being rebuilt around agents as first-class developers, and the storage layer is the part that had to change first, because everything else sits on top of it.

What it means if you just write code

If you are a developer who still writes code by hand, Rubicon will mostly show up as things not breaking. Faster CI, fewer merge traffic jams, and repositories that stay responsive while agents swarm them. The visible change will be gradual, since GitHub is building it underneath a running service.

The bigger shift is conceptual. GitHub is designing for a world where the majority of pushes come from agents, not people. That world is closer than most developers think, and Rubicon is the infrastructure bet that GitHub intends to be ready for it. Watch the preview reports over the next few months: if real teams report stable latency under agent load, the rebuild worked.

AS
News Cast Editorial Team

We cover AI, gadgets, startups and security — original reporting and explainers, written by humans, for humans. Got a tip? Reach us via the contact page.