Anthropic outgrew Git at 30,000 merges a day

Anthropic merged about 2,500 ⁰⁷ code changes a day in May 2026. By September it merged more than 30,000 ⁰⁸. That is 12 ²⁵ times the traffic in four months.

Git did not keep up. Seven engineers spent five months building Scion, an in-house replacement, and moved Anthropic's entire monorepo onto it. The company described the monorepo switch in an engineering post on October 13, 2026 ⁰⁶.

Every number in this piece comes from Anthropic's own post. No outside party has verified them. The lesson for everyone else is in what broke first.

The Human-Scale Ceiling in Source Control

Anthropic's post states the mechanism in one line: "The bottleneck, now, is all of the developer infrastructure built for human scale: source control, builds, and CI." Call it the human-scale ceiling, the point where tools sized for people typing start choking on agents shipping. Coding got cheap. The systems that move code around stayed priced for people.

Anthropic's figures map onto three parts of that ceiling. The first is load. Merges rose from about 2,500 ²⁵ to over 30,000 ⁰⁸ a day, and Anthropic says CI alone runs "on the order of 25 million ¹⁰ code checkouts per day."

The second is an old assumption. Anthropic says Git assumes clients work offline and need a complete copy of the repository. At Anthropic's size, that gets expensive fast. The company reports main-branch history of almost 100 GB ¹⁵ and a compressed working copy over 25 GB ¹¹.

The third is cost. Anthropic writes that fetched naively, its checkouts "would be over 600 PB ¹⁶ of git traffic per day." The math checks out: 25 million ²⁶ checkouts times 25 GB is 625 million GB, or about 625 PB.

One more division is worth doing. Take 25 million ²⁷ checkouts and divide by 30,000 merges, and you get roughly 833 CI checkouts per merged change. That ratio is our arithmetic, not Anthropic's. We think it is the most useful number in the whole post.

Inside Scion's Server-Side Bet

Picture a warehouse where every picker must photocopy the full inventory ledger before moving one box. With a handful of pickers, nobody notices. At 30,000 ⁰⁸ orders a day, the photocopier becomes the business. Scion puts a clerk at the counter instead: you ask for the rows you need, and the clerk keeps the books.

In system terms, Anthropic describes four moves. Step one, the server does the heavy work, so commits and merges run centrally. Rebases and history queries run there too. Step two, clients keep a bounded local cache. They fetch only the directories a sparse checkout requests, or materialize files on demand in a virtualized checkout. Step three, repository questions like blame become API calls, so a CI job can get an answer without cloning anything.

Step four puts regional caches between the jobs and the core. Anthropic says each cache tier runs above a 99% ¹⁷ hit rate. A hit rate is the share of requests answered from the cache without going back to the main store. At 99% ²⁸, the core server sees about one request in a hundred.

The data model changed too. Anthropic's post says Scion records each changeset as the files changed against its first parent, with content-addressed pointers to compressed file data. Content-addressed means a file is stored and found by a fingerprint of its contents, so identical files are kept once. That shape suits thousands of small agent patches landing on one huge repo.

Anthropic reports big results. Scion runs with a 200x ¹² smaller CPU and memory footprint than the old Git-mirroring setup. CI and research checkouts take seconds where they used to take minutes. Anthropic also says CI gained 25% ²⁴ more CPU capacity, and an internal code-review surface called Antique shipped alongside Scion.

Anthropic published no dollar figures, so run the translation on your own bill. Say your CI fleet costs $40,000 ²⁹ a month. A 25% capacity gain is worth about $10,000 a month in jobs you no longer have to buy. The 200x ¹² figure only compares Scion with Anthropic's previous mirrors, so read it as an internal before-and-after, not a benchmark.

The build story has a twist. Anthropic's post says an early internal snapshot of Claude Mythos Preview rewrote a handwritten prototype of the Git-to-Scion ingestion pipeline to run 4x ¹⁴ faster with 20x less memory. So the model that created the load helped build the fix. Our read is that this loop matters more than Scion itself: a seven-person team can now afford infrastructure that once needed a platform department.

The trade has a price. Git lets a developer commit and read history on a plane with no network. Scion assumes everyone stays online, so a central outage stops every agent at once. Superpower Daily's summary of the post says Anthropic stresses that Git remains suitable for most repositories.

Agents multiply checkouts faster than teams grow

Checkout Traffic Outgrows Headcount by 2027

The dated facts draw a short, steep line. There were about 2,500 ²⁵ merges a day in May 2026 and more than 30,000 ³⁰ in September. Anthropic published the fix on October 13. Nothing in the post suggests the human team grew anywhere near 12x over those four months.

The napkin projection is simple. Say the curve only doubles over the next four months. That gives 60,000 ³⁰ merges a day by early 2027. At Anthropic's current ratio of about 833 ²⁷ checkouts per merge, CI would need roughly 50 million ³¹ checkouts a day.

Whether the curve holds is unclear. Merged changes are a loose proxy for useful work. Agent patches can be tiny, and some get reverted a day later. A 12x ²⁵ jump in merges could mean far less than 12x more shipped value.

The mechanism still compounds, though. Each agent you add raises merges, and each merge triggers CI. Every CI run then pulls the repo. Headcount grows one hire at a time. Checkout traffic grows with every agent those hires supervise.

So the risk is asymmetric. Tracking your checkouts-per-merge ratio costs an afternoon. Finding the ceiling during a release week costs days of blocked CI and stalled agents. Superpower Daily reports that Anthropic does not expect many organizations to face the same pressure soon.

For most teams, we agree. We also think the first wall they hit will be CI traffic, long before Git's data model gives out. GitHub and GitLab already sell parts of the hosted layer where that server-side work would live. Sourcegraph sells pieces of it too. And Git's own partial clones and sparse checkouts can capture part of Scion's shape without a migration.

Measure Your Checkouts Per Merge This Week

The build is one script that tracks the ratio Anthropic's numbers expose: CI checkouts per merged change. You run it weekly and watch the line.

First, count merges per day on your main branch. Use this command: git log --first-parent main --since="2026-05-01" --format=%cd --date=short | sort | uniq -c. The first-parent flag follows only main's own line, so each merged or squashed change counts once. Paste the output into a sheet.

Then pull CI checkout counts for the same days. Most CI providers let you export job runs, and every job that checks out the repo counts as one. Divide checkouts by merges. Anthropic's figures put its ratio near 833 ²⁷, and yours shows how hard each new merge hits your fleet.

Next, price a single checkout. Time a fresh clone with time git clone followed by your repo URL. Then check its size with du -sh on the .git folder. Multiply size by daily checkouts for daily transfer, and multiply clone minutes by your runner's per-minute price for dollars.

After that, trim one job. Switch a single CI job to a partial clone with git clone --filter=blob:none. A partial clone downloads the commit history without file contents, then fetches each file the first time a command needs it. Add git sparse-checkout set with the folder you build, so the working copy holds only that path.

Something will break on the first run. Jobs that run blame, read version tags or diff against old commits may fail once they hit missing data. Log which job broke and why. Then either give that job full history or move it to a later stage.

Finally, schedule the script to run every Monday and chart merges against checkouts per merge. If the ratio climbs while your headcount holds still, agents are pushing you toward the ceiling Anthropic hit in September. Rerun the trim on the next most expensive job, and keep the sheet.