The development

What happened

AI coding agents generate Git pushes and repeated context-fetch reads as they change software. GitHub's Rubicon storage design targets that growing workload by separating durable objects from disposable serving replicas and ordering updates per reference. The 5 November disclosure reports internal agent-scale tests; availability remains limited private preview, with broader rollout planned for early 2027.

Key facts

  • Git packs live in Azure Blob Storage; references and the pack manifest live in a replicated Ref-CAS log. Serving replicas keep rebuildable caches rather than authoritative local state.

  • Packs become durable before reference transactions commit. Separate compaction workers merge packs, checking that the replacement preserves the input objects.

  • GitHub reports internal tests with 5,000 concurrent agents and 7,200 pushes per second to one repository. These are vendor test results, not a general production throughput guarantee.

  • GitHub says agent-driven development amplifies repository traffic: an agent pushes a change, then other agents fetch it for context. Rubicon addresses this AI coding workflow's storage load; it is not itself an AI model.

The context

Why it matters

The AI connection is infrastructure supporting many coding agents that create and share code, rather than a new model capability. Separating large immutable objects from small reference updates targets their different scaling bottlenecks; same-reference writes still contend. Readers should distinguish vendor tests and a private architectural preview from future APIs or proof that their repository already uses Rubicon.