Anthropic replaced git. 30,000 merges a day broke it
In May, Anthropic merged about 2,500 code changes a day. By September it was over 30,000. The post that explains what happened next says coding is no longer the bottleneck at Anthropic; source control, builds and CI built for human scale are. So a team of seven built a new version control system called Scion, plus a code review tool called Antique, code search and a source control IAM layer, and moved the whole monorepo off git in five months.
The numbers behind the decision: main branch history is close to 100 GB, a compressed working copy is over 25 GB, and CI alone performs on the order of 25 million checkouts a day. Fetched naively that is 600 PB of git traffic daily. Anthropic was running several hundred internal git mirrors to survive it. The insight was that git's whole design, every clone is a full offline server, was the wrong fit for a client base that is almost entirely agents and jobs that are always online. Scion flips it: the server does history, blame, commits, merges and rebases; the client keeps a bounded cache and a sparse or virtualized checkout; everything is exposed as an API so CI jobs often need no checkout at all; commits are append-only changesets whose cost scales with the diff, not the repo. The result is a footprint 200 times smaller than the mirror fleet in CPU and memory, checkouts in seconds instead of minutes, and a 25 percent CPU capacity win in CI.
Two details matter more than the architecture. First, Claude built most of it. An early internal snapshot of Claude Mythos Preview was asked to rewrite the hand-written git-to-Scion ingestion pipeline with no further instructions and produced a version that used 20 times less memory and ran 4 times faster while preserving the implicit invariants. Every engineer on the team now rates Claude as better at writing code than themselves, and the post says their job was very often to tell Claude no: agents violated core constraints to satisfy a pattern, and adversarial review made them over-engineer against hypothetical concerns. Reviewers now check intent, not lines. Second, the post admits a merge queue does not make sense at agent scale. Anthropic could not replicate a global repository lock, so it moved to post-merge and pre-deploy testing and lets Claude resolve the resulting conflicts faster than humans can find them. Component health is watched by Claudes they call superfans, one per subsystem, reading telemetry in a Slack channel.
This is the clearest datapoint yet for the thread we keep calling harness-not-the-model. The model did not get better in May; the infrastructure under it got rebuilt and the merge rate went up 10x. Anthropic is explicit that Scion is not a git replacement for anyone else and will be hard to open source, which means the first agent-native toolchain is also proprietary. Dex Horthy's read, that your VCS becomes part of production once no human is in the loop, got 56,000 impressions, and the summary posts went past 70,000. The HN thread sat at 402 points.
Links: https://www.anthropic.com/engineering/redesigning-source-control , https://news.ycombinator.com/item?id=50073633 , https://x.com/dexhorthy/status/2110080282279719187
← Back to all articles
The numbers behind the decision: main branch history is close to 100 GB, a compressed working copy is over 25 GB, and CI alone performs on the order of 25 million checkouts a day. Fetched naively that is 600 PB of git traffic daily. Anthropic was running several hundred internal git mirrors to survive it. The insight was that git's whole design, every clone is a full offline server, was the wrong fit for a client base that is almost entirely agents and jobs that are always online. Scion flips it: the server does history, blame, commits, merges and rebases; the client keeps a bounded cache and a sparse or virtualized checkout; everything is exposed as an API so CI jobs often need no checkout at all; commits are append-only changesets whose cost scales with the diff, not the repo. The result is a footprint 200 times smaller than the mirror fleet in CPU and memory, checkouts in seconds instead of minutes, and a 25 percent CPU capacity win in CI.
Two details matter more than the architecture. First, Claude built most of it. An early internal snapshot of Claude Mythos Preview was asked to rewrite the hand-written git-to-Scion ingestion pipeline with no further instructions and produced a version that used 20 times less memory and ran 4 times faster while preserving the implicit invariants. Every engineer on the team now rates Claude as better at writing code than themselves, and the post says their job was very often to tell Claude no: agents violated core constraints to satisfy a pattern, and adversarial review made them over-engineer against hypothetical concerns. Reviewers now check intent, not lines. Second, the post admits a merge queue does not make sense at agent scale. Anthropic could not replicate a global repository lock, so it moved to post-merge and pre-deploy testing and lets Claude resolve the resulting conflicts faster than humans can find them. Component health is watched by Claudes they call superfans, one per subsystem, reading telemetry in a Slack channel.
This is the clearest datapoint yet for the thread we keep calling harness-not-the-model. The model did not get better in May; the infrastructure under it got rebuilt and the merge rate went up 10x. Anthropic is explicit that Scion is not a git replacement for anyone else and will be hard to open source, which means the first agent-native toolchain is also proprietary. Dex Horthy's read, that your VCS becomes part of production once no human is in the loop, got 56,000 impressions, and the summary posts went past 70,000. The HN thread sat at 402 points.
Links: https://www.anthropic.com/engineering/redesigning-source-control , https://news.ycombinator.com/item?id=50073633 , https://x.com/dexhorthy/status/2110080282279719187
Comments