Anthropic replaced git after agents broke it

In May, Anthropic merged about 2,500 code changes a day. By September, it was merging more than 30,000. The jump, the company says, came largely from agents writing code. So Anthropic spent five months replacing git.

The new system is called Scion. Anthropic described it in an engineering blog post on October 13, and the numbers explain the decision better than any architecture diagram. The main branch of its monorepo weighs nearly 100 GB. A compressed working copy sits above 25 GB. CI alone pulls roughly 25 million checkouts a day, which fetched naively would be over 600 petabytes of git traffic per day.

Git was built for humans who work offline. That is the whole problem. As Anthropic put it, the properties that make git excellent for distributed, offline work were actively working against the company as the repo grew. Nearly all of its git clients are agents and CI jobs that are online by definition. Keeping a full local copy of history on every client, and syncing it constantly, stopped being an asset and became the bottleneck.

Scion flips the model. It is a server-centric, always-online system: the server answers questions about history, creates commits, and performs merges and rebases, while the client keeps a bounded cache and fetches only what it needs. The unit of work is a "changeset," an entry in a write-ahead log. Checkouts are sparse by design and can be virtualized, so a job only downloads the files it actually reads. Anthropic claims the new setup runs at a 200-fold smaller CPU and memory footprint than the old git-mirror farm, with regional caches hitting above 99 percent, and that CI checkouts are two orders of magnitude faster. Seconds, not minutes.

There is a feedback loop in this story worth sitting with. A seven-person team built Scion in five months, and Claude helped build the thing that manages Claude's output. The engineers asked Claude to rewrite an early Git-to-Scion ingestion pipeline with no further instructions, and the model produced a version that used 20 times less memory and ran four times as fast. The same post is candid about the other side of that loop: "Very often, we found that the role of the engineers was to tell Claude 'no.'" Agents, the engineers write, will sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place. The humans became reviewers. The reviewers became infrastructure.

Now the part that matters to everyone else. Anthropic is explicit that Scion is not for you. It calls the system "by no means a suitable replacement for git for general use," and there will be no open-source release. The company also hedges on timing, though the hedge points one way: it contends its own scaling problems will become commonplace in other organizations over the next few months. Read that as a forecast, not a product pitch.

The forecast is plausible because the pattern keeps repeating. Yesterday's Harvard and Jellyfish study found coding agents produce 30 percent more code and zero more finished software, because review time rose 49 percent. The bottleneck moved from writing code to reviewing it. Anthropic's post says the bottleneck then moved again, from reviewing code to the infrastructure that stores, versions, and tests it. GitHub itself attributed availability problems in April to the exponential growth in pull requests and commits. Three different chokepoints, one cause: the volume of machine-written code keeps finding the next human-scale system and overwhelming it.

What should an engineering leader do with this? Not install Scion, obviously. The useful signal is the design direction. Scion assumes clients are always online, pushes work to the server, virtualizes checkouts, and exposes everything through an API so tools can act without a local repo. Those are the properties an agent-scale codebase wants. If you run coding agents at any serious volume, your next infrastructure conversation is probably about sparse checkouts, CI caching, and whether your review tooling assumes a human pace. Anthropic's bet is that version control, builds, and CI all need the same treatment source control just got. The last two years suggest the bet is right.

What remains unknown: whether the 30,000-a-day figure keeps climbing at this rate (Anthropic says it sees no reason it would slow), how much of that volume is agent-authored versus human-authored (the post does not publish the split), and whether the Git community takes Anthropic's advice to add server-side improvements and better negotiation protocols for agent-scale workloads. None of the performance figures have been independently reproduced. They are Anthropic's measurements of Anthropic's system.

Related: the Harvard/Jellyfish review bottleneck is the companion story. The two pieces together describe where the agent-coding stack actually strains: not generation, but everything downstream of it.