BuildAlso reported elsewhere2 publishers2 min readPublished
Anthropic built Scion after agents pushed daily merges from 2,500 to over 30,000
Anthropic built a new source-control system, Scion, after daily merges rose from about 2,500 in May 2026 to over 30,000 in September. A dev.to account of its October 13 post says review, CI and merge tooling built for human contributors became the bottleneck.
The Engineer · Build desk
Key figures card. Anthropic says Scion's infrastructure footprint is approximately 200 times smaller than its earlier Git mirroring approach in CPU and memory terms. Anthropic reports cache hit rates above 99%. A seven-person team built Scion and Antique in five months, per Anthropic.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Smaller CPU and memory footprint than old Git mirroring | 200 times, approx. | 16 | |
| Cache hit rate, above | 99% | 15 | |
| People on the team that built Scion and Antique | 7 people | in five months, per Anthropic | 17 |
What happened
- Anthropic's October 13, 2026 engineering post presents Scion with Antique, a code-review surface, plus code search, access management and large-scale landing and merge tooling.
- Scion models work as changesets instead of Git commits and runs commits and merges on the server, so clients do not push in the conventional sense.
- The server side is distributed and horizontally scalable, and no single server holds a fully materialized copy of the repository.
- Anthropic says a seven-person team built Scion and Antique in five months.
Why it matters
- constraint Anthropic cautions that Scion is not a drop-in Git replacement, and copying it elsewhere could mean reworking the whole development lifecycle, not swapping a version-control server.
- exposure Anthropic's post leans on per-directory access controls and semantic limits on review approvals, because at agent volume a loose permission or a mechanical approval gets used far more often.
- decision A team adopting coding agents has to decide which of five places to instrument first: CI capacity, review queues, landing workflows, access controls, or search and dependency tooling.
Between May and September the volume grew roughly twelvefold [13]. The dev.to write-up says the constraint was the combination of repository size, access patterns and the rise in changes [4]. Its argument about where output stalls is general: if review queues, CI jobs, repository synchronization or merge operations cannot keep pace, they become the practical limit on AI-assisted engineering output [7].
The reported results come from the caching layer. Anthropic built it to separate failure domains [14]. The company reports cache hit rates above 99% and CI and job speedups from minutes to seconds in relevant workflows [15].
Treat those as measurements of one monorepo under one access pattern. They would transfer to another repository only if its size and access pattern were close to Anthropic's, since the write-up names that combination as the constraint [4].
The footprint claim needs the same care. Anthropic says Scion's infrastructure is about 200 times smaller than its earlier Git mirroring approach in CPU and memory, contributing to an estimated 25% CPU capacity gain [16]. A footprint ratio and a capacity gain are different quantities, and the write-up does not give the baseline for either.
The dev.to piece also points to a symmetry. According to Anthropic, Claude's coding capabilities repeatedly accelerated both the migration and the follow-on workflows, so AI raised the pressure on existing developer infrastructure and helped the team build the replacement [18].
What to watch
- Whether Anthropic publishes the baselines behind the 200-fold footprint and 25% CPU capacity figures.
- Whether daily merge volume stays near 30,000 or keeps climbing after September.
- Whether other teams with agent-heavy monorepos report where their queues saturate first: CI, review or landing.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption10
- Hype gap+10
- Incentives60
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic's daily merge volume rose from roughly 2,500 changes in May 2026 to more than 30,000 in September, and the company redesigned its internal source-control infrastructure afterward.
- [2]
Anthropic's engineering post on redesigning source control, published October 13, 2026, describes Scion alongside Antique, a code-review surface, plus code search, identity and access management, and large-scale landing and merge tooling.
- [3]
Source control, builds and continuous integration systems built for human-scale activity became bottlenecks as agents accelerated the rate at which code could be produced and submitted.
- [4]
Anthropic's constraint was the combination of repository size, access patterns and a sharp rise in code changes, not merely the size of a monorepo.
- [5]
Scion uses changesets rather than traditional Git commits, and moves commits and merges to the server, so clients do not need to push changes in the conventional sense.
- [6]
Scion's server architecture is distributed and horizontally scalable, with no single server holding a fully materialized copy of the repository.
- [7]
If code review queues, CI jobs, repository synchronization or merge operations cannot keep pace with code generation, those systems become the practical limit on AI-assisted engineering output.
- [8]
Anthropic explicitly cautions that Scion is not a drop-in replacement for Git, and most software teams do not face 30,000 daily merges.
- [9]
Reproducing Anthropic's approach elsewhere could require reworking the entire development lifecycle, not just replacing a version-control server.
- [10]
Anthropic's post points to monorepo security, per-directory access controls and semantic constraints in review approvals.
- [11]
Higher throughput can increase the consequences of weak permissions or overly mechanical approval processes.
- [12]
The dev.to piece tells teams adopting AI coding tools to watch CI capacity and test duration, code review and approval queues, merge conflicts and landing workflows, repository access controls and permissions, and search, dependency understanding and developer feedback loops.
- [13]
Daily merge volume grew roughly twelvefold between May and September 2026.
- [14]
Anthropic built a high-performance caching layer to separate failure domains.
- [15]
Anthropic reports cache hit rates above 99% and CI and job speed improvements from minutes to seconds in relevant workflows.
- [16]
Anthropic says Scion's infrastructure footprint is approximately 200 times smaller than its earlier Git mirroring approach in CPU and memory terms, contributing to an estimated 25% CPU capacity gain.
- [17]
Anthropic says a seven-person team built Scion and Antique in five months.
- [18]
According to Anthropic, Claude's coding capabilities repeatedly accelerated both the migration and follow-on workflows; AI increased pressure on existing developer infrastructure while also helping the team build infrastructure meant to handle it.
Sources
2 independent publishers whose own reporting we read for this story.
- anthropic.comAnthropic details Scion, a git replacement for its monorepo, built by seven engineers with Claude in five months, as daily merges rose from 2.5K in May to 30K+
1 article · October 12, 2026
- dev.toAnthropic’s Scion Shows Why AI Agents Are Forcing a Rethink of Source Control
1 article · October 13, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.