Anthropic didn’t hit a coding bottleneck — it hit a post-coding bottleneck. Its Claude models write so much code that Git couldn’t keep up. By September 2026, the company was merging over 30,000 code changes per day, up from around 2,500 in May. That surge wasn’t driven by human engineers scaling up.
It came from Claude itself, generating, iterating, and submitting code at volume no traditional development workflow was built to absorb. As Anthropic put it in its internal blog post: 'We don’t expect this rate of growth to slow down anytime soon. The latent demand for merging code is practically infinite, and every bottleneck we remove immediately leads to a further increase in the amount of code we produce.' The problem wasn’t writing code.
It was storing it, reviewing it, searching it, checking it out. At scale. Their main branch weighs about 100 GB. A compressed working copy exceeds 25 GB.
And they perform roughly 25 million code checkouts daily just for CI, before counting developer environments, agent sandboxes, and research jobs. Naive Git fetches would generate over 600 PB of traffic per day. Even with Git’s partial-clone features, operations proved unwieldy. So Anthropic built Scion: a centralized monorepo system, designed from the ground up for AI agents.
Unlike Git, Scion shifts work to the server. Clients are thin, queries run via API, and synchronization is minimal. Its unit of work is the 'changeset': an entry in a write-ahead log that lists changed files and points to compressed content in blob storage, not tree hashes. Development began in January 2026.
When Claude Mythos Preview became available internally, engineers prompted it to rewrite an early Git-to-Scion ingestion pipeline, and got a version using 20 times less memory and running four times faster. Engineers still matter. But their role has shifted: 'Very often, we found that the role of the engineers was to tell Claude “no,”' Anthropic said.
'Agents will sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place.' Scion won’t be open-sourced. It’s not meant for general use. But Anthropic says the pressure it solved. AI-generated code overwhelming legacy tooling, will spread fast.
And the fix won’t be off-the-shelf.
CDFiled by The Computing Desk
Laptops, desktops, processors, graphics cards and the operating systems that run on them.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.