Redesigning source control at Anthropic

Vince Thirty thousand merges a day is such a deranged number that the interesting part isn’t Scion. It’s that Anthropic is basically saying code generation stopped being the bottleneck.

Ava Right.

Ava And I think that’s the actual article. Not, you know, 'behold our custom version control system.' Their claim is that git, builds, and C I were built for human-scale output, but their workload is now mostly always-online agents and jobs.

Vince Yeah, and that is the most Exploring Next thing imaginable. We spend almost a year doing this show, and the answer is still: the boring plumbing is where reality shows up.

Ava Okay, wow.

Vince Also, you know I love this kind of post. Not because I want anyone rebuilding source control for fun. Because it names the real product boundary. If code gets cheap, merge throughput, review, permissions, and receipts become the thing you actually buy.

Ava Mm-hm.

Ava The numbers are what make it hard to wave away. They say the main branch history alone is almost one hundred gigabytes, a compressed working copy is over twenty-five gigabytes, and they do on the order of twenty-five million checkouts per day for C I. Naively, that would be over six hundred petabytes of git traffic per day. At that point, 'just use shallow clone' is not a serious answer.

Vince And they did the normal hacks first. Mirrors, sparse checkouts, promisor remotes, repo seeds in blob storage. So this isn’t a vanity rewrite. It reads like they kept patching human-era infrastructure until the patch layer became the product.

Ava Right, right.

Ava Technically, the cleanest move is the assumption flip. Git treats every client like a potential full server with offline needs. Scion says most clients are online agents, so let the server do history, blame, merge, rebase, commit creation. The client keeps a bounded cache and a working copy if it even needs one.

Vince Which is also why the API part matters more than the V C S branding. If a job can ask for a diff, merge base, blame, or path history without hauling a repo onto disk, you’re not just speeding up developers. You’re changing what counts as a software operation.

Ava Exactly.

Vince And yes, I can hear you loading the skepticism cannon already.

Ava Oh, come on.

Ava My skepticism is actually pretty mild here. This is one of those cases where the workload really is weird enough to justify weird infrastructure. They even say most orgs do not have this problem yet. I buy that. Where I’d push back is if anyone reads this as 'git is obsolete now.' No. Anthropic is closer to Google-than-normal on monorepo scale.

Vince Yeah, no, you’re completely right. The practical audience is not 'every startup, go fork git.' It’s teams leaning hard into coding agents who should maybe notice that their next fire is review queues, CI saturation, repo access patterns, and who approved what.

Ava Sure.

Ava The data model is interesting too. Their unit is a changeset in a write-ahead log. Instead of storing a full tree pointer like git, it stores changed files relative to the first parent plus content-addressed blob pointers. Then blame, path history, copy detection, file search, all that comes from derived data computed recursively from parents. That’s a very systems-person answer to 'how do I keep commits cheap but still answer expensive questions later.'

Vince It’s such a good example of your favorite thing, honestly. Don’t make every operation pay for every future possibility. Make the write cheap, then materialize the extra structures where they earn their keep.

Ava No way.

Ava You did just say my favorite thing back to me, but yes. Also I appreciate that they discourage refs except for main or release branches, because global shared pointers become coordination pain at scale. That detail makes this feel real, not brochureware.

Vince Tiny side note, Antique is an incredible name for the review surface. Scion and Antique sounds like somebody let infrastructure engineers name a fantasy bookstore.

Ava Okay, that’s genuinely funny.

Ava But it also fits the bigger pattern we keep tripping over. Same as the receipts thing from that code-cleanup story. Once models can generate absurd amounts of change, the scarce resource is not output. It’s legible evidence, safe landing, and cheap queries against reality.

Vince That’s the part that’ll stick for me. Not 'Anthropic replaced git.' More like, if your agents are always online, maybe your tools should stop pretending they’re lone laptops on a plane. I’d put seventy-thirty that within three months we see at least one major devtools company ship a very explicit server-side repo query layer for agent workflows, not just faster clone marketing.

Ava I’d take lower, maybe sixty-forty, just because shipping a real query layer is harder than tweeting the architecture diagram. But yeah, Vince, that’s the next obvious pressure point.

Vince All right. Episode ten eighty-one, and once again the unglamorous part wins. I cannot believe this is how we spend a Thursday.