# GitHub's Rubicon redesign targets thousands of simultaneous agent pushes to one repository GitHub is testing a Git storage redesign that separates immutable objects from ordered reference updates. Its 7,200-pushes-per-second figure is an internal test, not a generally available service guarantee. Project Rubicon uses object storage, disposable replicas and a reference transaction service to accommodate agent-driven parallel Git writes; access remains limited. - Status: Active - Published: 2026-11-10T18:49:33+13:00 - Updated: 2026-11-10T18:49:33+13:00 - Categories: Web Development, Cloud & Infrastructure, Deployment & DevOps, Developer Tools - Tags: Coding agents, Developer tools, Git, GitHub - Canonical HTML: https://beyondthe.news/dossiers/github-rubicon-git-storage-agent-concurrent-pushes-2026 ## What changed On November 5, 2026 GitHub detailed Project Rubicon, a new Git storage architecture in limited private preview. Immutable packfiles and indexes live in Azure Blob Storage, while a replicated Ref-CAS service orders the small set of reference and pack-manifest changes that need consensus. Disposable virtual replicas handle traffic, with background compaction separated from request serving. GitHub says an internal test supported 5,000 concurrent agents and 7,200 pushes per second to a single repository. ## Why it matters Many parallel coding agents create sustained, branch-heavy write loads unlike traditional human commits. Separating content upload from small atomic reference updates can reduce contention and let repositories scale without changing Git commands. For platform teams, this is an architectural signal, not an immediate migration: availability is limited and GitHub's test results are not independent production measurements. ## Why the old write path was a bottleneck GitHub says typical high-traffic repositories historically saw bursts of dozens of writes per second, whereas agents may sustain far more. A previous blob-based reference write-ahead log serialised updates, and splitting the log across blobs raised read overhead. The redesigned service commits reference changes using compare-and-swap transactions, locking only the affected references. ## How Rubicon splits the work Incoming Git packs are indexed and uploaded to Azure Blob Storage. A Ref-CAS transaction makes the pack visible and changes the affected branch references. Virtual replicas can be added for hot repositories; their disks are rebuildable caches rather than authoritative state. Compaction runs separately under shard leases to control the number of packs. ## Performance claim and rollout boundary GitHub reports 5,000 concurrent agents and 7,200 pushes per second to one repository in internal tests, while serving low-latency reads. The public engineering account does not provide an independently reproducible benchmark suite, cost comparison, or production latency distribution. The architecture is in limited private preview, with wider rollout planned for early 2027 rather than general availability now. ## What developers may eventually notice Ordinary push, fetch and branch workflows are intended to remain compatible. GitHub also describes possible agent-oriented APIs for reading selected files without full clones, atomic multi-file commits and event streams. These are roadmap items, not currently guaranteed features of the preview. ## Independent context Stack Archive's November 6 technical summary independently interprets the architecture and its potential branch-per-agent use case, but it is based on GitHub's engineering disclosure, not an independent load test. OpenAI's separately documented Pando design shows that agent-driven Git storage is a broader infrastructure problem, though its performance figures are not directly comparable. ## Key details - GitHub engineering article dated November 5, 2026; Project Rubicon in limited private preview. - Azure Blob Storage holds immutable Git packs and supporting indexes. - Ref-CAS coordinates reference and manifest updates; disposable replicas serve traffic. - GitHub internal test: 5,000 agents, 7,200 pushes/second on one repository; independently unverified. - Git protocol and ordinary developer commands are intended to remain unchanged. - Broader rollout expected early 2027; agent-specific APIs are future plans. ## Builder takeaways - Do not plan capacity around 7,200 pushes per second as a production service guarantee. - Keep parallel agent changes on separate branches to minimise reference contention and merge collisions. - For self-hosted Git systems, examine object/reference separation and cache/compaction costs before copying this architecture. - Watch preview eligibility, rollout details and actual production reliability before changing agent orchestration. ## What to watch - Broader Rubicon rollout and public availability dates. - Production latency, consistency and cost evidence. - Actual availability and constraints of file-read, atomic commit and event-stream APIs. - Branch contention, merge bottlenecks and repository protections at high agent concurrency. ## Uncertainties - The performance result is vendor-produced, not independently replicated. - GitHub has not published representative production cost or long-term failure data for Rubicon. - Limited private preview does not imply availability to all repositories. - Future agent-specific APIs remain proposals or roadmap statements. ## Sources - [Inside GitHub's new Git storage architecture](https://github.blog/engineering/architecture-optimization/inside-githubs-new-git-storage-architecture/) — GitHub Engineering · primary engineering report · 2026-11-05T00:00:00+13:00. Architecture, internal benchmark and rollout statements. - [GitHub Rubicon: Git Storage Built for 5,000 Concurrent Agents](https://stack-archive.com/blog/github-rubicon-git-storage-architecture-agents-2026/) — Stack Archive · independent technical interpretation · 2026-11-06T00:00:00+13:00. Derivative architectural analysis; not independent benchmark. - [Scaling Git at OpenAI](https://openai.com/index/scaling-git-at-openai/) — OpenAI · primary comparison · 2026-11-02T00:00:00+13:00. Separate architecture illustrating agent-driven Git infrastructure pressures; metrics not directly comparable.