Anthropic rebuilt source control for agent-scale merges

If your CI or monorepo is straining under agent-written commits, read Anthropic's Oct 13 engineering post. The company says merges to its monorepo rose from about 2,500 per day in May 2026 to over 30,000 per day in September, so it built Scion, an in-house version control system, and fully migrated off git in five months. Scion is not a product you can download.

What changed#

Anthropic published "Redesigning source control at Anthropic" on Oct 13, 2026 (@AnthropicAI posted at 12:36 PM ET). Coding agents made human-scale git the bottleneck: main-branch history alone nearly 100 GB, compressed working copy over 25 GB, on the order of 25 million CI checkouts per day (before developer, agent, and research jobs). Naive fetch math would be over 600 PB of git traffic per day. At peak they ran several hundred internal git mirrors.

Scion flips git's offline-first design: clients are always online; the server creates commits, merges, rebases, blame, and history; clients keep a bounded cache and sparse or virtualized working copies. A "changeset" stores files changed relative to the first parent (not a full tree). Anthropic also shipped Antique (code review), code search, and source-control IAM in the same window. A seven-person team built Scion and Antique. Infrastructure footprint is now about 200x smaller in CPU and memory than the old git mirroring setup; CI and research checkouts are "two orders of magnitude faster (in seconds rather than minutes)"; CI alone saw a 25% CPU capacity win.

Why it matters#

You cannot switch to Scion today. The post is a forward map for anyone running swarms of coding agents against a large monorepo: expect mirror farms, seed lag, and checkout time to fail before model quality does. Anthropic says this class of problem will become more common over the next six to 12 months and that it wants to share ideas with GitHub, GitLab, and the git community rather than productize Scion.

How to use it#

Measure merges per day, CI checkout duration, and number of git mirrors or seed snapshots you already run. Anthropic's pain started when agent merge volume outgrew human-scale tooling.
Prefer sparse checkouts and server-side "give me the diff / merge base / blame" APIs where your forge offers them, instead of full clones for every agent job.
Plan for post-merge and pre-deploy testing if a global merge-queue lock cannot keep up when hundreds of changes land while tests run (Anthropic's read of agent-scale CI).
Keep humans on intent review: Anthropic engineers say they still review every change request, but they now examine intent more than every line, and often tell Claude "no" when agents violate core constraints.

Our monorepo averages N merges/day and CI checkout takes T seconds. Using Anthropic's Oct 13 Scion post as context (not as a product we can install), list the top five git-scale failure modes we should watch as coding agents increase merge volume, and one metric or forge feature for each.

Good to know#

Anthropic is explicit: "Scion is by no means a suitable replacement for git for general use" and mainly fits huge monorepos with tens of thousands of engineers or prolific agent swarms.
Adopting Scion would mean migrating code review, search, and CI; internal tools are "highly bespoke."
Claude Mythos Preview rewrote their git-to-Scion ingestion pipeline to use 20x less memory and run 4x faster with no further instructions, per the post.
"Superfans": Slack-channel Claudes tasked to watch one component's health via telemetry.
Watch next: forge announcements of server-side RPCs, better sparse/shallow composition, or agent-oriented merge queues; Anthropic cites Piper and Sapling as prior art.

Sources#

https://www.anthropic.com/engineering/redesigning-source-control (published Oct 13, 2026, checked 2026-10-14 5:57 AM ET)
https://x.com/AnthropicAI/status/2110047081414987957 (created 2026-10-13 12:36 PM ET / 16:36 UTC, checked 5:56 AM ET)

Last checked: 2026-10-14