A team can adopt coding agents to ship faster and discover that code generation is no longer the limiting step. The constraint moves into the path between a finished change and an integrated one: checkout, diff, review, testing and merge. Anthropic says that is what happened inside its monorepo.
In an engineering post published October 13, the company introduced Scion, an internal source control system that replaced Git for that repository. Anthropic reports that its merge volume rose from roughly 2,500 changes per day in May to more than 30,000 in September. The migration took five months.
Those figures come from Anthropic and have not been independently validated. The case is still useful because it exposes a concrete architectural shift. When code production grows by an order of magnitude, tools built around every developer holding a complete copy begin moving the cost somewhere else.
What Scion gives up to gain scale
Git lets every clone operate as a self-contained repository. History, diffs and core operations remain available offline. That is the right tradeoff for most projects. In Anthropic's workload, however, the main branch history was approaching 100 GB, a compressed working copy exceeded 25 GB, and CI alone performed about 25 million checkouts per day.
Scion starts with the opposite assumption: agents and jobs are always online. Its servers create commits, perform merges and rebases, answer history queries and maintain derived structures for blame and path search. The client keeps a bounded local cache and materializes only requested files. Sparse and virtualized checkouts are the default operating model rather than a compatibility layer.
Anthropic says every cache tier achieves a hit rate above 99%, the new infrastructure uses 200 times less CPU and memory than its previous Git mirror fleet, and CI and research checkouts now take seconds instead of minutes. The company attributes a 25% CI CPU capacity gain to the change. These are internal results from an unusual workload, not a general benchmark.
The improvement also depends on an explicit tradeoff. The client can do little without the server. Anthropic exchanges part of Git's offline autonomy and resilience for a centralized architecture that is observable and horizontally scalable. That may be sensible for a fleet of always-online agents. It is a much less attractive bargain for a developer who needs to work without a network.
The lesson is not to replace Git
Anthropic explicitly says Scion is not a general-purpose Git replacement. It is internal, requires its own review, search and CI stack, and was built for a huge monorepo with token spending that few organizations would accept. Google Piper and Meta Sapling already showed that extreme repositories demand different choices. The new variable is the change rate produced by agents.
There is a revealing tension in the company's account. Anthropic credits Claude with accelerating Scion's development and says a seven-person team built Scion, its Antique review system and several related tools in five months. The same post says engineers still review every change and often need to tell agents “no” when they violate core constraints or add complexity in response to hypothetical risks.
MnzAI Labs' assessment is that this detail matters more than a “Git versus Scion” debate. Automation expanded both implementation capacity and the need to define invariants. Human work shifts toward architecture decisions, safety boundaries and the intent behind a change. Without that judgment, speed simply delivers conflicts and technical debt to a faster queue.
Teams far below 30,000 daily merges can still use the case to decide what to measure:
- how much time agents spend downloading and preparing a repository;
- where review, CI and merge queues form;
- which operations truly require a complete local copy.
The answer may be better caching, sparse checkout and richer server APIs built around Git. Anthropic itself points the ecosystem in that direction. The strategic signal is broader: if agents increase code production as promised, developer infrastructure needs to be measured like a production system. The bottleneck does not disappear. It changes address.