Cursor has built Continuity, a Git storage system that achieved more than 300 pushes per second in synthetic tests by replacing traditional replica coordination with an S3-backed write-ahead log as the single source of truth. The company announced the architecture in a blog post published in September 2026 and reported linear read scaling with as many as 100 replicas. The design addresses both massive repositories with intensive continuous integration demands and the proliferation of smaller repositories generated by coding agents.
In synthetic tests using Cursor's everysphere monorepo, the system delivered up to 120 pushes per second with S3 Standard storage and exceeded 300 pushes per second when using S3 Express One Zone. At the higher throughput level, Git compaction operations emerged as the limiting factor rather than storage performance. Cursor reports that conditional S3 reads used to verify replica state averaged under 10 milliseconds. The company says tested pushes were linearizable and saved to external storage before acknowledgment, while clones stayed fully consistent throughout. The architecture enables monorepos to support hundreds of replicas for continuous integration workloads, and idle repositories can be created on demand rather than maintained continuously.
Cursor engineer Vicent Martí characterized local NVMe repositories as warm caches instead of authoritative copies in the new model. The system stores pushed data in S3 and logs the matching reference update in the write-ahead log, with pushes confirmed only after required data has been saved, delivering durability before acknowledgment. According to the report, rendezvous hashing picks preferred nodes while atomic compare and swap operations on S3 let any server accept a push. Cursor also batches operations to minimize the effect of S3 PUT latency on throughput, and UDP gossip spreads WAL updates across nodes. The company notes that lost gossip doesn't compromise correctness because S3 remains the source of truth.
The storage approach has drawn comparisons to database systems from engineers reviewing the architecture. Maksim Al Dandan, a senior software engineer, described the method as treating Git storage like a database and pointed to push consistency, force-push transactions, and point-in-time recovery as questions applicable to the model. Casey Lee, CTO at Liatrio and a former AWS engineer, highlighted the architectural shift and noted that Cursor's published performance numbers haven't been independently verified. The Continuity model changes how replication and compaction scale by moving the consistency boundary from the replica layer to durable object storage. Instead of synchronously coordinating additional replicas, Continuity allows local repositories to converge independently on state stored in S3, exchanging replica coordination for storage validation, asynchronous propagation, and increased bandwidth. Only the primary server performs Git compaction, with replicas downloading the resulting packs from S3, which trades bandwidth consumption for CPU resources.
The architecture represents a fundamental rethinking of how distributed version control scales when storage infrastructure can guarantee consistency without tight replica coupling. Organizations managing large repository fleets may find the trade-offs particularly relevant as agent-generated repositories multiply and traditional coordination overhead compounds.