Anthropic replaced Git with a homegrown version control system called Scion after AI agents drove the company’s daily code merges from 2,500 to over 30,000 — in four months. That 12x surge made Git’s distributed architecture not just slow but structurally unworkable. So a seven-person team built a replacement in five months, migrated the entire monorepo, and are now publicly warning the rest of the industry that it’s coming for them too.
The Numbers That Broke Git #
In May 2026, Anthropic averaged around 2,500 merged code changes per day. By September, that number had climbed past 30,000. The main branch history alone hit nearly 100 GB. The compressed working copy exceeded 25 GB. CI jobs were running 25 million checkouts per day — a figure that would have required 600+ petabytes of daily Git traffic if handled naively.
Git’s response to these numbers is essentially: use shallow clones, partial clones, and mirrors. Anthropic tried all of that. The operational overhead of maintaining hundreds of internal mirrors became unsustainable, and Git’s offline-first design kept surfacing fundamental limits: packfile compression latency, broken blame and history queries on partial clones, and merge coordination that doesn’t scale to 30,000 operations a day.
What Scion Actually Is #
Scion inverts Git’s design philosophy. Where Git assumes offline-capable clients that hold full or partial local copies, Scion assumes clients are always online and fetches only what they need. Commits, rebases, and merges all happen server-side. Clients maintain bounded caches — not full repositories.
The core unit is a changeset: an entry in a write-ahead log that stores the list of changed files relative to its parent, plus content-addressed pointers to compressed blobs in object storage. No packfiles. No cross-server coordination. No locking. Because all data is content-addressed, any server in the three-tier topology (scion-servers, regional scion-caches, per-node scion-daemons) can cache objects without coordinating with others.
Checkouts are sparse by default. Virtualized checkouts can materialize files entirely on-demand. The system discourages feature branches — only main and release branches exist, by design. Branches are global state requiring server consensus, and global state is a scaling bottleneck at this velocity.
The Results #
The seven-person team shipped Scion plus Antique — a companion code review tool — and completed the full monorepo migration within five months. Performance outcomes were substantial: 200x reduction in CPU and memory footprint versus the previous Git mirror fleet, checkout times dropping from minutes to seconds, 25% of CI CPU capacity freed, and cache hit rates above 99% across all tiers.
Claude helped build its own infrastructure replacement. The model rewrote the git-to-Scion ingestion pipeline, achieving a 20x memory reduction and 4x speed improvement. It also independently identified an algorithmic optimization to ancestry queries. The irony that Anthropic’s AI agents created the scale problem and then helped engineer the solution is hard to miss.
What Anthropic Is Not Saying #
Anthropic is explicit that Scion will not be open-sourced, and equally explicit that it is “by no means a suitable replacement for git for general use.” The system only makes sense for enormous monorepos supporting tens of thousands of engineers or dense AI agent swarms — most organizations are nowhere near this scale today.
What they are saying: other organizations will hit this wall within 6-12 months as AI agent deployment intensifies. Rather than releasing Scion, Anthropic is sharing the architectural thinking — novel git server architectures, comprehensive server-side RPCs standardized across forges, better composition of sparse and shallow features — and asking GitHub, GitLab, and the Git project to evolve before the pressure arrives.
The Broader Signal #
Scion did not arrive in isolation. OpenAI acquired Pierre Computer, a developer tooling startup, on November 5. Meta acquired Conductor, a multi-agent coding orchestration platform, on November 4. Three major AI labs, in the same week, all moving to control the developer infrastructure layer. The O’Reilly Radar for November 2026 flagged source control innovation as a leading trend, noting that “coding is no longer the bottleneck; infrastructure is.”
Git was designed in 2005 for the Linux kernel — humans, working intermittently, across time zones. Scion was designed for Claude: always online, never sleeping, capable of merging hundreds of changes per hour. These are categorically different users, and the same tooling was never going to serve both indefinitely.
For most developers, nothing changes today. Git is not going anywhere. But the engineering decisions that GitHub, GitLab, and the Git project make over the next 12 months — in response to pressure from customers running exactly this kind of AI-heavy workflow — will matter a great deal. Anthropic just told you what they’ll be pressuring them to build. Read the full Anthropic engineering post — it’s unusually candid about both the problem and the limits of the solution.