Anthropic Rebuilt Git (yes really)
Thursday, Oct 15 · 48 min · about 10 min to read
Theo-brain watched all 48 min
Theo reads Anthropic's post on Scion, an in-house replacement for Git (the standard tool that tracks code changes). Anthropic built it as daily merges rose from about 2,500 to over 30,000. Theo's lesson: tools built for human-speed work no longer fit agent-speed work, and AI has made rebuilding cheap. So rebuild what slows you down, and make bad changes easy to undo rather than hard to make.
Skim itSkim the briefing for the main points, and watch only if you want to hear Theo's opinions in his own words.
Parts
Why Git is showing its age
Git was designed for a world where code was expensive and humans reviewed everything. Theo says it no longer fits how work is moving.
variables. Seriously though, Git is not the right primitive for the way that things are moving right now. I'll be
Anthropic's merge explosion
Anthropic went from about 2,500 merged changes a day in May to over 30,000 in September.
things like that. In May of 2026, Anthropic averaged around 2,500 code changes merged a day. As of September, it's over 30,000 a day. We don't expect
Others have rebuilt source control too
Meta built its own system because its huge checkouts didn't work well with Git. Theo says you shouldn't force the standard tool into a workflow it doesn't fit.
that's fine. You shouldn't have to force the industry standard into your workflow if it doesn't fit. And it is nice seeing more companies challenge this cuz Git is not perfect and it certainly is not perfect for agents. So let's learn more
Scion's design principles
Scion's server does the heavy work, and commits must not lock each other out, so hundreds of agents can commit at once.
piece, commits must be cheap, incremental, and lock-free. Lock-free is particularly important. One person committing shouldn't block somebody else, or more realistically, 500 agents committing shouldn't block another 500 agents. In Git, you commit locally and
How the data is stored
Code lives in cloud storage as files. Scion is essentially a database pointing at them, not a full copy of the repo.
but kind of crazy to see the like the direct callout that the source code does not exist in the system and when you request the source, it has to go fetch that from blob store and it's effectively just a database pointing at those blobs. Seems like all the metadata
Layers of caching
Each cache tier serves over 99% of requests, and the setup needs 200x less computing than Anthropic's old Git mirrors.
possible. Each tier of this system maintains a cache hit rate of over 99% and it doesn't require long-lived state. So, it's easily scaled up if needed. Their infrastructure footprint is 200x smaller than their previous Git mirroring infra in terms of CPU and memory requirements. That's insane.
Building bespoke tools with AI
Theo argues switching costs have collapsed, so writing your own alternative is now easier than staying stuck with a vendor.
Historically, people adopt a technology, a solution, a vendor, and then they kind of are stuck there cuz switching off is more expensive than whatever cost they're incurring staying. Now, it's easier to write your own alternative and move to it.
Engineers' new role and the speedups
The post's benchmarks show everyday commands running many times faster, for example pulls 25x and status 207x.
system instead of Git. Pulls are 25x faster, checkouts are 14x faster, rebase is 5x faster, status is 207x faster, commits are only 30% faster, disappointing, but diffs are 6x faster.
Security and operations
Scion can lock down individual folders, so sensitive code stays protected inside one big shared repo.
supports rich access controls on repositories, as well as the ability to apply permissions to individual directories in order to protect the most sensitive code without losing the benefits of the mono repo. Look who was
Rethinking old assumptions
Anthropic dropped the one-at-a-time merge queue and test after merging instead, which Theo welcomes.
haven't been able to effectively replicate the industry standard of a global repo lock on the merge queue. So, instead, they've invested in post-merge and pre-deploy automated testing. Look
Not a Git replacement for everyone
Anthropic won't release Scion. Theo predicts falling token prices could let a company build something similar for a few thousand dollars.
The speed that these prices are going down is insane. So, it is absolutely possible by this time next year you'd be able to write your own similar thing to Scion, your own alternative to Git for your company, for just a few thousand dollars in tokens. They also call out a
Tap any time to jump the video there.
Key takeaways
Scion's core insight: Git assumes everyone keeps a full offline copy of the code. Nearly all of Anthropic's users are agents that are always online, so Scion keeps the work on a central server.
partial clones. Taking a step back, we noticed that, like most companies, we were using Git as a centralized system with one authoritative copy. But, nearly all of our clients were agents in jobs that are online by definition. This is a
The payoff was large. Automated test runs and research jobs now get their code in seconds instead of minutes. Test infrastructure alone freed up 25% of its computing capacity.
CI and research jobs are able to check out code two orders of magnitude faster, seconds rather than minutes, which frees up a sizable amount of capacity across the fleet. CI alone had a 25% CPU capacity win. Neither clients nor caches
A team of just seven built Scion and Antique, their code review tool. When Anthropic's new model, Mythos preview, was asked to rewrite a hand-built prototype, its version used 20x less memory and ran 4x faster.
small teams is going to become more common. Can confirm. A team of seven people built both Scion and Antique, their code review tool, which they could
They asked it to rewrite the whole thing and with no further instructions, it produced a version that consumed 20x less memory and was four times faster
According to the post, the engineers' job was often to tell Claude no, because agents sometimes break core design rules to follow a pattern. Theo jokes they must be using the same Claude as everyone else.
knew they needed to achieve. Very often, they found that the role of engineers was to tell Claude no. Agents will sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place. Okay, maybe they are using the same Claude that we are.
Anthropic itself says Scion is not a general Git replacement. It only makes sense for enormous repos with tens of thousands of engineers or swarms of agents.
specify this is in no means a suitable replacement for Git in general use. The concepts of outlined here really only make sense for truly enormous mono repos with either tens of thousands of engineers or swarms of agents prolifically writing code. Plus,
Three skills here are new to you.
De-risking the merge button New
- What it is
- Make merging safe on two fronts. Code should be very likely fine by the time a person approves it. And a mistake that slips through should do little damage, because it goes to a few users first and is easy to revert (undo).
- Why Theo cares
- This is how he merged 50+ pull requests (proposed code changes) in about four hours one night. He timed it before a nightly build so he could test it and pull the build if it was bad.
- How to try it
- Ask your agent to set up releases that go to a limited group of users first.
- Make reverting a bad change a quick, routine step.
- Batch changes into a regular build you personally test before it reaches everyone.
also the other side, where if you merge something you shouldn't have, the blast radius and the damage that it can do is reduced massively by systems to make it easy to test the code once it goes out with a limited set of users, and revert the things if they do actually end up
because I wanted to have as much in as possible for the nightly, so before I went to bed, I could download it, test it, make sure all was good, fix anything that was broken, and revoke the build if it was bad, and then go to bed smiling
Replace a vendor with your own build New
- What it is
- Have agents build a narrow, custom version of a paid tool that fits only your needs.
- Why Theo cares
- He had agents copy all his analytics data from a provider into his own ClickHouse database (a fast analytics store) on R2 (Cloudflare file storage). It was more than 10x cheaper, easier for his agents to use, and he did it part-time alongside his main work.
- How to try it
- Pick one vendor whose cost or limits annoy you.
- Ask your agent to mirror and backfill your data from it into storage you control.
- Run both side by side until you trust the copy.
It's insane. I personally threw together a solution to clone and mirror and backfill all of our analytics data that I have on one particular analytics provider so that I have it directly on my own managed ClickHouse on top of R2. Ended up being more than 10x cheaper and way easier for my agents with. And I did this part-time with a handful of threads on the side while I was working on my
Claude "super fans" for monitoring New
- What it is
- Anthropic gives Claude Slack channels, each focused obsessively on the health of one system component, with access to its logs and metrics.
- Why Theo cares
- Anthropic says these catch subtle problems that normal alerts miss. Theo, who mocks the name, admits it seems cool, though he calls it burning compute for fun.
- How to try it
- Give an agent read access to one service's logs and metrics.
- Ask it to watch for anything unusual and post findings to one channel.
- Keep your normal alerts running alongside it.
For example, they have Slack channels where they tasked Claude with being obsessively focused on the health of single components. They call these super fans. Also, they they refer to them as Clauds, plural. That hurts me a bit.
detect subtle and novel issues that make those alerts seem one-dimensional by comparison. That does actually seem kind of cool. Everything earlier I said where they were reducing compute with the programmatic thing, this is the opposite. This is them burning compute for fun.
Advice
If you aren't shipping this fast, it's either because mistakes could kill people or because you haven't set up the tooling that makes fast shipping less risky. Set it up.
now. But this is reality. If you're not moving this fast yet, you are either in like a place where you absolutely can't because people will die, or you just don't realize how powerful these things are and you haven't taken the time to set up the tooling to make the stuff less risky. So, let's dive in a bit
Drop the expert's reflex to trust only proven tools. Use what you know about where tools break to reinvent the ones you rely on for your own needs.
times. You got to get over that. Things are changing too fast. You have to flip the way you think about this almost like upside down where you have this knowledge of what does and doesn't work, which leads you to picking certain tools. Now you need to do the opposite.
Hot takes
Being able to undo a bad change matters more than piling on layers of process to prevent one.
His reasoning: No matter how strict your process is, you will ship something broken, so fixing it fast and limiting the damage is what counts.
you guys have no idea how long I've been harassed for thinking it's more important to be able to undo a bad change than to add thousands of layers in front of the changes to make it slightly harder to make a bad change. So
something when it breaks. Because no matter how strict your processes are, you will ship something broken. I don't care who the [ __ ] you are. If you
Anthropic writing code to sort reviews by type, instead of just asking Claude, is what he's most bearish on. He wonders if they're short on compute.
His reasoning: You could simply hand Claude the code and ask, so building it programmatically suggests they can't spare the compute.
an answer. This actually might be the thing I'm most bearish on in this whole thing. If they are compute constrained enough that they're not running Claude to summarize and filter their diffs, and they feel like they need to make programmatic solutions for this instead, maybe they are out of compute. Maybe we
Anyone who says Gemini is in a good state right now can't be trusted as an engineer.
His reasoning: Gemini 4 read his whole database schema and then queried columns that don't exist, while Claude was doing algorithmic optimizations.
video about Gemini 4 Argon. Gemini 4 was hallucinating columns in my SQLite tables. After reading the file that is my whole schema, it would then fire a query for columns that don't exist.
Anybody who says Gemini is in a good state right now is not trustable as an engineer.
Many of his problems with Git really come from GitHub, which he calls slow and miserable.
His reasoning: External estimates put its downtime as high as 7%. He has had releases blocked by failing automated checks and merges blocked by pages that wouldn't load.
real with y'all. A lot of my problems with Git are coming from GitHub because it's just a slow and miserable platform. External estimates are measuring as much as 7% downtime across different things within GitHub. I can't tell you how many times I've had releases blocked because actions were failing or I haven't been able to merge an important change because the page wasn't loading. The
How Theo thought about this
He credits Anthropic's claim of better reliability because it matches what he has seen himself. Claude Code and the API have been dependable lately, and he has recently had more outages from OpenAI and Codex.
Claude was down. The website goes down decently often, but like Claude code and the Claude APIs, those are pretty reliable lately. Funny enough, I think I've had more problems with OpenAI and Codex outages than I have with Claude recently, which is unbelievable considering how big the gap was before.
Sign in to mark what you've watched and keep notes.
Have a question about this one? Ask about this video →
Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.
