It’s 11:40 on a Tuesday night and I’m watching three agents work a single repo. One is rewriting a retry wrapper, one is patching tests it broke twenty minutes ago, and the third is doing something to the Dockerfile that I will regret. My terminal is stuck on a clone. Not a long one — maybe ninety seconds — but I’m sitting there watching a progress bar while three processes that never sleep wait on me and my local disk.
That ninety seconds is the whole story. It’s fine when a human triggers it a few times a day. It’s a tax when a fleet of always-on jobs triggers it constantly.
Anthropic hit that wall at a scale most of us can only simulate with a cron job and bad judgment, and their answer was to stop patching around Git and build something else. The system is called Scion, and they’ve already migrated to it internally.
The number that forced the rewrite
In May 2026, Anthropic’s monorepo averaged around 2,500 code changes merged per day. By September 2026, that figure was over 30,000 per day. Same company, same repo, roughly four months apart, twelve times the merge volume.
Sit with that as an infrastructure problem rather than a productivity brag. Writing code stopped being the constraint. What’s left is everything downstream — source control, builds, CI — and those systems were designed around the assumption that a human types for a while, commits, and goes to lunch. Remove the lunch. Remove the human. Multiply the actors. The assumptions collapse in a specific and predictable order, and version control is near the front of the line.
What Scion actually changes
Anthropic describes Scion as built for always-online agents and jobs, which is a design brief more than a feature list. Three pieces stand out:
- Incremental changesets. Work arrives as deltas rather than as repeated full-history operations. When thousands of jobs are each touching a handful of files, the unit of work should be the handful of files.
- Scalable object caches. Objects get served from shared caches instead of being re-fetched into every worker’s private copy of everything.
- Sparse or virtualized checkouts. A job materializes the slice of the tree it needs and nothing else.
The reported results are substantially faster checkouts and a smaller infrastructure footprint than Git. Both of those matter, but the second one is the quieter win. Faster checkouts you feel. A smaller footprint is what keeps the bill from scaling linearly with your agent count.
The monitoring detail I keep thinking about
Anthropic also deployed cloud Tag instances and Slack channels to watch the health of specific components. That’s not a headline item, but it’s the part that tells you they expect things to break in ways nobody will be staring at a dashboard to catch. Per-component health, piped into the place humans already are. When your writers are machines and your reviewers are machines, the humans become operators, and operators need alerts that find them.
What this means if you’re building bots at normal scale
You are not going to write your own version control system. Neither am I. But the failure mode Anthropic designed around is the same one that shows up in a side project with four agents and a GitHub Actions budget.
- Stop giving every worker a full clone. Sparse checkouts and shallow clones exist in Git today. Use them before you need them.
- Cache the objects, not the repo. A shared cache layer between your agents and your remote is the cheapest version of what Scion does natively.
- Measure merge throughput as a system metric. If you only track whether agents produce good diffs, you’ll miss the moment the pipeline becomes the constraint.
- Instrument per component. One green checkmark across a pipeline hides which piece is quietly degrading. Anthropic split theirs out for a reason.
- Treat CI as the next wall. Source control was first because it’s upstream of everything. Builds and test runs are standing right behind it.
The useful takeaway isn’t that Git is finished. It’s that Git was tuned for a workflow where the expensive, slow, careful part was a person thinking. We’ve automated the thinking part and left the plumbing untouched, and the plumbing is where the next round of engineering work lives.
Mine is still a ninety-second clone. Anthropic’s was 30,000 merges a day. Same shape, different order of magnitude — and they got there in four months.
🕒 Published: