Anthropic left Git after 30,000 daily merges changed the rules

Anthropic built an in-house version-control system called Scion and fully migrated its monorepo in five months. The company says merged changes rose from about 2,500 per day in May 2026 to more than 30,000 in September.

Thirty thousand daily merges no longer fit Git's operating model

The main branch had almost 100 GB of history and a compressed working copy exceeded 25 GB. Anthropic says CI performed roughly 25 million checkouts per day. A naive Git transfer would amount to more than 600 PB each day, so the company ran several hundred internal mirrors at peak.

Scion reverses a basic Git assumption. Clients are always online, while the server creates commits, handles merges and rebases and answers history queries. A client keeps a bounded cache and downloads only requested trees and blobs. A seven-person team also built the Antique review system and completed the migration within five months.

Agents moved the bottleneck from code generation to infrastructure

Anthropic describes an environment where most clients are online agents and automated jobs. A full offline repository copy has less value for them than it does for a developer with a laptop. A centralized service can therefore expose cheap APIs for diff, blame, bisect or revert without a local checkout.

The company reports that Scion's infrastructure uses 200 times less CPU and memory than its previous Git-mirroring fleet. CI and research checkouts became two orders of magnitude faster, moving from minutes to seconds, while CI alone gained 25% in CPU capacity. The lesson for other teams is narrower: measure whether their real constraint is code creation, review, testing or data movement.

Scion solves Anthropic's extreme case, not the ordinary company repository

The figures are self-reported and Scion is not offered as a public product. Anthropic says the design makes sense for an enormous monorepo with tens of thousands of engineers or swarms of agents, and acknowledges that its internal token spending is unrealistic for other businesses. Migration also required custom review, search and CI systems.

Centralization sacrifices offline work and makes server availability more important. Anthropic reports cache hit rates above 99% at every tier, but provides no independent evaluation or operating-cost breakdown. Performance on one unusual workload is not a universal verdict on Git.

The next pressure point will be review, testing and deployment

The useful signal will be whether teams outside Anthropic see similar merge growth and whether GitHub, GitLab or open-source tools adopt richer server queries and better sparse workflows without demanding a full migration. Anthropic expects this class of problem to spread over the next 6 to 12 months.

Review has already changed inside the company. Humans still inspect every change, but focus more on intent than every line. If agents keep accelerating code production, the advantage will not come from a faster commit. It will come from stopping a bad intention before cheap code multiplies it.

Lilith's verdict

Anthropic built a highway for 30,000 merges a day. The scarce part is no longer the asphalt but the person at the barrier who spots a change heading the wrong way.