Anthropic, Cursor, GitHub, and OpenAI are evolving their version control systems to adapt to agentic scale, a usage scenario where software agents work in parallel on the same code. Each is moving away, in its own way, from Git’s purely decentralized model, where each node hosts a complete, autonomous copy of the history.
- Anthropic remplace Git par Scion, dont le serveur gère les opérations lourdes.
- Cursor garde le format Git et ajoute Continuity, un journal d’écriture sur S3.
- OpenAI conçoit Pando, qui adopte la cohérence causale et réduit la latence des push de près de 80 %.
- GitHub limite le quorum et le protocole 3PC à chaque référence, visant un débit d’écriture jusqu’à 35 fois supérieur.
Anthropic Replaces Git with Scion
On its internal monorepo, a single repository containing all its code, Anthropic went from about 2,500 merges per day in May 2026 to over 30,000 by September. The main history is around 100 GB, and nearly 25 million daily fetches are performed for continuous integration (CI). According to the company, using Git in a conventional way would have required transferring over 600 PB of data per day. Running hundreds of mirrors and snapshots of the repository on object storage had become unmanageable.
Anthropic therefore replaced Git with an in-house system called Scion, developed in 5 months. The server handles heavy operations like merges, rebases, and commit creation, while the local client is limited to a lightweight layer that manages a bounded cache and materializes files on disk on demand.
A Scion changeset only stores the list of files changed relative to the first parent, unlike a Git commit which points to a complete tree of objects. Scion also eliminates packfiles, the compressed archives Git uses. According to Anthropic, the infrastructure consumes 200 times fewer compute and memory resources than its old Git fleet, and CI has gained 25% CPU capacity.
Cursor Relies on an S3 Write-Ahead Log
Cursor retained Git’s native storage format and tools, on local NVMe disks. To overcome the limits of Spokes, GitHub’s historical replication architecture, the company designed an alternative engine named Continuity, applied to its monorepo and a fleet of micro-repositories.
This architecture relies on Write-Ahead Logs (WAL), which store multiple objects sequentially in a single file. Each push forms a distinct entry in this log, and no push is acknowledged until it is persisted to S3. An index file, updated via CAS (compare-and-swap, an atomic conditional update operation), orders each transaction. The replicas on NVMe disks are merely caches, rehydrated from the log, and any node can receive a write.
Continuity achieves 120 pushes per second on S3 Standard and over 300 with S3 Express One Zone. Beyond that, the intrinsic limit is Git’s on-disk compaction speed. For reads, the system showed linear scaling up to 100 replicas, with no impact on writes.
OpenAI and GitHub Adjust Their Architectures
In two years, the volume of commits arriving on the master branch of OpenAI’s monorepo increased roughly 25-fold, and version control reliability had fallen below 90%. In 10 weeks, the company developed Pando. This backend abandons providing a consistent state of all branches at a given point in time, adopting causal consistency instead: a client respects the history of a branch’s author, without a global order relative to unrelated branches. Servers reuse immutable packfiles and store data on object storage, with no local decompression during pushes. OpenAI states that Pando reduced median push latency by nearly 80% and can deliver nearly 1 PB of Git data per day with a “handful” of CPU cores.
For its part, GitHub recorded 7.38 billion commits in September 2026, for 3.35 billion pushes, a volume that more than quintupled in a year. Its architecture, currently being deployed to production, limits quorum and the 3PC protocol to individual references, rather than the entire repository. Internal tests showed a 3 to 4 times higher write throughput on concurrent repositories. Authoritative data is entrusted to Azure object storage, and GitHub announces a potential write throughput gain of up to 35x, with autonomous scaling of the read layer.
