Cursor, GitHub and Anthropic Each Describe New Source-Control Infrastructure for Agent-Driven Change Volume, While Zed Adds a Thread Layer Above Git

Background

Git is built as a decentralized system. In Anthropic’s description, every clone is a complete, self-contained repository that can compute diffs, read history and blame, and check out any revision while offline, and Anthropic says this is the right design for the vast majority of repositories. Hosting platforms nonetheless run Git as a centralized service. GitHub’s engineering blog says every repository is stored by a system called Spokes, which keeps a full copy on the local disks of several fileservers, five by default, and uses a three-phase commit protocol with a quorum so that CI, the web interface and API clients see a consistent state. GitHub adds that adding replicas to absorb read load makes writes slower.

Several sources describe that arrangement coming under pressure from software agents. The Register reported on 28 October that GitHub recorded 7.4 billion code commits in September 2026, five times the number a year earlier, and that the surge led to outages; it also reported that the site’s 90-day availability hit a low of about 84 percent on May 7, according to data reconstructed from GitHub’s status page. The New Stack reported that GitHub had said in August it was seeing 2.9 billion commits per month, up from 1.4 billion in April. Anthropic reports a similar curve inside its own monorepo: about 2,500 code changes merged per day in May 2026 and over 30,000 per day as of September.

This article compares four public accounts, published between August and October 2026, of how vendors are changing source-control infrastructure in response. The figures in the vendor posts are the companies’ own, and the independent outlets cited here largely relay and contextualise those posts rather than test them.

Current situation

Cursor: an object-storage log behind Git

The Register reported that Cursor principal systems engineer Vicent Martí wrote a post on how Cursor arrived at the architecture for Origin, its Git-based repository service, which is powered by an internal engine called Continuity. According to The Register, pushes are uploaded into S3 in a write-ahead log that captures changes as immutable objects, and are simultaneously written to a local reference copy of the repository. Martí wrote that agents have made the situation worse, as quoted by The Register: “More code, more PRs, more CI runs.” The Register concluded that it remained to be seen how the approach performs as Origin moves into production. The Herald covered the Continuity design and the Origin launch in August.

GitHub: separating durability from serving

In an October 6 post, GitHub says agentic development is driving “the next architectural shift” and that it is rebuilding its Git infrastructure while the service keeps running. The post gives its own workload figures: pushes grew 4.9x year over year, from 0.69 billion to 3.35 billion per month, pull request merges grew to nearly 4x their volume a year earlier, and GitHub Actions ran 3.26 billion times in September. It lists “Merges contend on one reference” among the pressure points, noting that merge queues funnel work onto a single ref.

GitHub’s stated design keeps authoritative repository data in Azure Blob Storage, with read capacity coming from lightweight workers that cache data. In internal benchmarks, GitHub says it has delivered up to 35 times higher write throughput. In its Universe announcement on 28 October, Jay Parikh, Microsoft’s EVP of CoreAI, wrote, as quoted by The New Stack, that “Agents are changing the volume and shape of software development.” The New Stack reported that GitHub says it has added four million CPU cores and 160 petabytes of high-speed storage so far this year, while The Register described the same figures as being on GitHub’s infrastructure roadmap, alongside a migration to a multi-region deployment in Azure that Parikh said is expected to complete early next year.

Anthropic: Scion replaces Git for one monorepo

On 13 October Anthropic published a post on Scion, an in-house version control system to which it says it fully migrated its monorepo from Git in five months. Anthropic says its main-branch history is almost 100 GB, that it runs on the order of 25 million CI checkouts per day, and that it had operated several hundred internal Git mirrors at its peak. In Scion, per the same post, the server creates commits, merges and rebases, source-control queries are exposed through APIs so jobs often need no local repository, and the client holds only a bounded object cache. Anthropic reports a 200x smaller CPU and memory footprint than its previous mirror infrastructure and checkouts that are two orders of magnitude faster. InfoQ labelled these “company-reported results for its internal workload, not general benchmarks against Git.”

Anthropic calls Scion “by no means a suitable replacement for git for general use” and The Register reported that there will be no open source release. The Register placed Scion alongside earlier large-repository systems, naming Google’s Piper, Microsoft’s Virtual File System for Git and Meta’s Sapling. Anthropic also credits Sapling’s server component, Mononoke, with pioneering many of Scion’s techniques. On merge queues, Anthropic wrote in the same post that it had not been able to replicate the industry-standard global repository lock on its merge queue and had instead invested in post-merge and pre-deploy automated testing, accepting that concurrent changes can conflict, which it says Claude can typically fix faster than humans can identify them.

Zed: a thread layer above Git

The New Stack reported in mid-September that Zed had launched Delta in public beta, a collaborative environment in which developers and agents work in shared threads instead of pull requests. Delta currently works with Git repositories; its DeltaDB layer records edit-level activity between commits. Zed says it has disabled pull requests on Delta’s own repository and that 33 developers have landed 570 changes to main without them. The report also states that Cursor launched Origin in August and that GitLab is working on a “next-generation source-code management” project called Project Switch, in private beta. Zed’s chief executive Nathan Sobo is quoted saying that “GitHub has an established social component that will take longer to replace,” and that Zed is moving more thoughtfully with its public repository, which has hundreds of monthly contributors. The Herald covered the launch earlier.

Implications

The following points are drawn from comparing the sources above; they are not predictions.

Where authoritative state lives. The three infrastructure accounts differ in scope but share a direction. Per The Register, Cursor treats a write-ahead log in S3 as the source of truth. GitHub says authoritative data will sit in Azure Blob Storage with cache workers in front, and that reads and writes then scale independently. Anthropic stores content-addressed file contents in blob storage and serves them through tiered caches. In each description, a set of full local copies is not what makes the data durable.

Whether Git stays. Cursor and GitHub describe changes beneath Git workflows; GitHub says its new infrastructure is designed to support the same branching, review, merge and history workflows at much higher volumes. Anthropic left Git for its own monorepo but recommends that Git platforms add comprehensive server-side RPCs and improved fetch protocols, and The New Stack reported that Anthropic wants to partner with GitHub, GitLab and the open-source community. Zed keeps Git storage for now, according to The New Stack.

Merge queues are treated differently. GitHub names single-ref merge contention as a load to be engineered around, while Anthropic says it has not been able to replicate a global lock on its merge queue. Both accounts describe the same pressure from opposite sides, and the sources do not say which approach suits other organisations.

How far the lessons travel. Anthropic says it anticipates this class of problems becoming more commonplace over the next six to 12 months, while also saying its own workload is unusual and that its ability to spend tokens internally far exceeds what any other business would realistically want to spend. InfoQ’s takeaway for architects was to reconsider where work happens as workloads change, noting that Scion depends on server availability and that the required integration is substantial.

What remains unknown

None of the performance figures cited here has been independently reproduced in the sources reviewed. GitHub’s 35x figure comes from internal benchmarks, Anthropic’s figures are described by InfoQ as company-reported, and Cursor’s design is, in The Register’s words, still to be tested in production. GitHub’s migration is described as still in progress. The sources do not say whether Delta’s thread model or Project Switch will gain adoption, and they do not give a timetable for any of the Git-platform changes Anthropic suggests.