Cursor published the storage system behind Origin, and named agent repo sprawl as the design driver

A day after launching Origin, its git forge, Cursor published the engineering account of what sits underneath it. Martí walks through why hosting git is hard, why GitHub's distributed-filesystem attempts failed, and how Spokes, the consensus-based replication system built at GitHub around 2013 and since copied across the industry, works and where it now breaks. Cursor's replacement is called Continuity: a write-ahead log kept in S3-compatible object storage as the source of truth, with ordinary git repositories on local NVMe treated as a warm cache, no routing tables and no relational database. Pushes are serialized by an atomic compare-and-swap on S3; replicas learn about them over UDP gossip and verify against S3 with a conditional GET, which Martí reports returns a 304 in under 10ms on average. First-party figures: synthetic stress tests to 100 replicas with linear read scaling, sustained ingest of up to 120 pushes per second on S3 Standard and more than 300 per second on S3 Express One Zone, measured against Cursor's own monorepo.

Martí connects the storage design to coding agents through his case against three-phase commit. Spokes needs three replicas per repository whether the repository is busy or idle, and, he writes, "When agents work with Git repositories at scale, they often operate outside of a monorepo by creating vast numbers of small repositories, many of them throwaway, and most of them barely touched." Under Continuity a monorepo can spread across hundreds of replicas for CI load while millions of agent-created repositories run on one each, and an idle one is garbage collected off disk and rebuilt from the log on the next fetch. The post is a vendor making a case for its own product, and the numbers in it are self-reported with no independent reproduction.

Why it matters: Cursor has published the first technical basis anyone outside the company has for evaluating Origin, along with the claim that existing git hosting was architected for a repository population that agents have already invalidated. For a team weighing where its source of truth lives, the scaling argument is checkable in outline, since the failure modes of three-phase commit under many replicas are not controversial. The trust argument is not addressed here at all: Cursor closed a SpaceX acquisition four days ago, and this post says nothing about training use of hosted code, which was the question the Origin launch thread spent its time on.