Key details

  1. Engineering disclosure dated October 13, 2026; Anthropic says migration completed in five months.
  2. Company-reported daily merged changes rose from about 2,500 in May to more than 30,000 in September.
  3. Main-branch history nearly 100 GB; compressed working copy over 25 GB; approximately 25 million CI checkouts per day.
  4. Internal Scion and Antique systems built by seven engineers; not publicly available.
  5. Vendor-reported 200x smaller source-control CPU/memory footprint, 100x faster checkouts, and 25% CI CPU capacity gain.
  6. Always-online clients, server-side operations, immutable blobs, tiered caches and sparse/virtualised working copies are central design choices.

What builders should take away

  1. Before replacing Git, measure checkout volume, mirror costs, history-query latency and CI cache misses; most teams will not face Anthropic's scale.
  2. Use existing Git sparse checkout, partial clone and caching features where they solve the bottleneck; distinguish repository size from change throughput.
  3. Agent-heavy pipelines should expose cheap history, blame, diff and merge-base queries to tools without forcing every job to clone the repository.
  4. Invest in post-merge/pre-deploy validation and explicit invariants as automated change volume rises.
  5. Treat Anthropic's cost and performance numbers as one organisation's internal case study, not a guaranteed general benchmark.

What changed

In an October 13, 2026 engineering disclosure, Anthropic said it replaced Git for its main monorepo with Scion, a bespoke server-centric version control system, and migrated the repository in five months. Anthropic says daily merges rose from about 2,500 in May to more than 30,000 by September; its monorepo's main history approached 100 GB and CI alone requested roughly 25 million checkouts daily. The seven-person team also built Antique code review, code search and source-control access controls. Scion is internal infrastructure, not a public replacement for Git.

Why it matters

This is a concrete case of AI coding throughput shifting bottlenecks from code generation to repository, review and CI infrastructure. The transferable lesson is that always-online automated workers may need cheap server-side history and diff APIs, incremental commits and sparse or virtualised working copies rather than repeated full local clones. Most builders should not migrate away from Git: Anthropic explicitly says its architecture only makes sense for extreme monorepos and agent swarms, and its measured gains are self-reported.

Why Git mirrors stopped scaling

Anthropic reports nearly 100 GB of main-branch history, a compressed working copy above 25 GB, and around 25 million daily CI checkouts. The company used shallow clones, sparse checkouts, snapshot seeds and several hundred Git mirrors, but freshness and history queries became expensive as merge volume accelerated. Its clients were predominantly online agents and jobs, so the offline independence central to Git was less valuable for this particular workload.

Scion makes the server do the source-control work

Scion stores changesets as write-ahead-log entries with content-addressed file references. Servers handle merges, rebases, blame, diff, ancestry and commits; clients keep bounded caches and materialise files only when requested. The design minimises global refs and packfile operations. CI jobs can often query the source-control service directly without a working copy, while regional caches and node daemons fetch immutable blobs on demand.

The performance numbers need attribution

Anthropic reports a 200-fold reduction in CPU and memory infrastructure footprint compared with its former Git mirroring fleet, checkouts two orders of magnitude faster, and a 25% CI CPU-capacity gain. It also reports cache-hit rates above 99%. These are internal vendor measurements without published independent benchmarks or a like-for-like general Git comparison. They should not be treated as predictions for typical repositories.

AI changed the engineering process, not just the tools

Anthropic says a seven-person team built Scion, Antique and related services within five months with heavy Claude assistance. Engineers continued to review changes, often focusing on intent, invariants and architecture; they report correcting agents that pursued brute-force approaches or unnecessary complexity. The company shifted attention toward post-merge and pre-deploy tests because conventional global merge locks are difficult at its rate of change.

Not a general-purpose Git replacement

Scion is neither open-sourced nor offered as a product. Anthropic says it would require organisations to replace surrounding code review, search and CI systems and that its economics are atypical. The company proposes working with GitHub, GitLab and the wider ecosystem on more scalable Git-compatible server APIs and sparse workflows instead of asking everyone to adopt Scion.

What to watch next

  • Whether GitHub, GitLab or Git tooling adopts standardized server-side source-control RPCs or improved sparse protocols.
  • Independent studies of repository/CI bottlenecks under high-volume agent workloads.
  • Any public technical artefacts or open-source components Anthropic releases from Scion or Antique.
  • Further detail on correctness, access-control, incident rates and long-term operating costs.

Still unclear

  • All reported performance and throughput metrics originate from Anthropic and lack independent reproduction.
  • The share of merged changes authored autonomously by agents is not quantified in the announcement.
  • Scion's code, operating cost and migration trade-offs are not independently inspectable.
  • Applicability to ordinary repositories and smaller teams is limited.

Sources

Direct reading behind this dossier.

3 sources

Discussion

Discussion is reader-contributed. Comments are not part of the BTN dossier or its editorial evidence.

0 visible comments

Join the discussion

Keep comments useful and relevant. Reader contributions may be moderated and are not BTN editorial evidence.

Sign in to comment