30,000 AI-generated merges a day broke git. Anthropic built something else.
Jump to a section
See the story at a glance
In May 2026, Anthropic's codebase was receiving about 2,500 code merges per day. By September, that number was over 30,000. The increase — 12× in four months — came almost entirely from AI agents.
Git stopped working.
The main branch history hit 100 GB. A compressed working copy exceeded 25 GB. At 30,000 merges and 25 million CI checkouts per day, a naive git fetch would have generated over 600 petabytes of traffic daily. Anthropic was running workarounds — shallow clones, sparse clones, hundreds of internal git mirrors at peak — and the workarounds were becoming their own operational problem.
So seven engineers spent five months building Scion, an in-house version control system designed from scratch for agent-scale development. Anthropic published the technical details on October 13. It's a quietly remarkable post if you read it for what it implies.
What Scion actually is
The core design premise: clients are always online, and the server handles everything that scales badly locally. No full copy of the repo lives anywhere. No packfiles. Clients fetch exactly what they need. The basic storage unit is a changeset — changed files relative to the first parent, with content-addressed blob pointers — stored in a write-ahead log.
The tiered fetch system: a per-node daemon, a regional cache, and an authoritative backend. Each tier maintains over 99% cache hit rate. CI and research checkouts now take seconds instead of minutes — roughly two orders of magnitude faster. CI alone gained about 25% CPU capacity. Infrastructure CPU and memory dropped by about 200× compared to the git mirroring setup.
There's one number in the post that stopped me: a Claude model rewrote the git-to-Scion ingestion pipeline. The result used 20× less memory and ran 4× faster than what the engineers had written, while preserving the ordering invariants. Scion is a system built partly to handle AI-generated code, built partly by AI.
The engineering role question
The post includes something that reads like an observation from inside the change, not a feature announcement:
Engineers focused on design tradeoffs and often needed to reject AI-proposed patterns that violated system constraints. Reviews shifted toward the intent of changes rather than every line.
That's a description of a job that is not "read the diff carefully." It's "understand what this change is supposed to accomplish and whether the system will behave correctly." Those are different skills. The second one is harder in the specific way that writing has gotten easier and judgment has gotten relatively more valuable.
Anyways — the engineering-role question is its own conversation. The infrastructure question is what I want builders to hold onto from this post.
Source spread
- Anthropic Engineering — Redesigning source control at Anthropic — [builder]. The primary source. First-party post from the team that built Scion. All quantitative claims come from here; raw data is not released.
- Anthropic — Claude Fable 5.2 release notes — [hype]. Fable 5.2 release states that over 90% of changes merged into Anthropic's production codebase are now Claude-generated. The source control post explains the infrastructure that makes that number possible.
Pros & cons
What's real:
- The 12× merge volume growth is a production throughput measurement, not a demo. These are real merges into a real codebase during a real four-month period.
- Scion's design principles — online clients, server-side history, content-addressed blobs, sparse working copies — address real scaling problems any organization running agents at this volume will hit. The technical decisions are coherent.
- 200× less CPU and memory versus the git mirroring setup is an efficiency gain that gets noticed at real scale. That's not a margin-of-error improvement.
What deserves a side-eye:
- This is a first-party technical post. The 99% cache hit rate, 200× CPU reduction, and seconds-not-minutes claims are from the team that built the system. No independent audit exists.
- Scion is not being open-sourced. The design insights are publishable; the software isn't. Most builders can't replicate this without building it themselves — Anthropic explicitly says it would require replacing the whole dev toolchain.
- "Hopes git platforms will adopt server-side RPCs" is a reasonable but optimistic read of the ecosystem trajectory. Git's architecture is deeply entrenched. Plan for this being a multi-year adoption curve.
Samwise's take
What builders need to know
- Git will hit a wall before you expect it. If your engineering org runs agents at meaningful scale, merge volume grows faster than linear. Plan for infrastructure changes before you're in crisis-mitigation mode.
- The design principles are replicable, even if Scion isn't. Online clients, server-side history operations, sparse working copies — these patterns can inform how you configure GitHub or GitLab today, even before you need a full replacement.
- Code review is changing. At agent commit rates, reviewing every line doesn't scale. Reviewing intent and architecture does. This requires different tooling and different reviewer judgment than current code review workflows assume.
- These numbers are plausible but not independently audited. If your infrastructure roadmap depends on any of these figures, test your own assumptions before committing.
- Watch how git platforms respond. GitHub and GitLab are tracking this problem. Expect server-side merge processing and sparse-checkout optimization to become competitive differentiators in developer tooling over the next 12-18 months.
Further reading
- Anthropic Engineering — Redesigning source control at Anthropic — the primary source
- Anthropic — Claude Fable 5.2 release notes — the 90% AI-merge number that this infrastructure exists to support
- Anthropic Engineering — Building strong safety evaluations — today's companion post on the operational practices Anthropic applies to AI-generated systems
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
robots.txt gets a shared AI training opt-out. Finally, one that might work.
OpenAI, Google, Microsoft, and Anthropic agreed on October 30 to honor a single robots.txt token — __usage_ai-training — for opting out of AI training. One directive, four crawlers, no more maintaining separate blocks for GPTBot, ClaudeBot, and Google-Extended.
Anthropic's 280-million-install standard just became everybody's. That's the right call.
Anthropic donated Agent Skills — the SKILL.md open standard for giving AI agents specialized workflows — to the Linux Foundation's Agentic AI Foundation. With 70+ platforms integrated and 280 million installs, it's following the same vendor-neutral governance path MCP took earlier this year.
Microsoft built a model that doesn't talk. That's the point.
Microsoft launched Decision-1 on October 9, post-trained from Alibaba's Qwen3.5-9B, that returns probability scores for a fixed option set rather than generating text. At $0.042/M input with output tokens free and 35× faster than GPT-6 Sol at P50 latency, it reframes what the routing layer of an AI system should look like.