When AI Writes Too Much Code, Anthropic Rebuilds Source Control
monorepo
One very large place holding many code projects.
Scion
Anthropic’s system for saving and finding code changes.
sparse checkout
A way to keep only requested parts of a codebase.
What happened
Anthropic, the AI company that makes Claude, described a major internal change in its engineering post. The company built Scion, an in-house source-control system. It fully moved its huge monorepo from Git to Scion in five months. Anthropic also built a code-review tool called Antique, code search, and new access controls. The post focuses mainly on Scion.
Anthropic said its daily merged changes grew from about 2,500 in May 2026 to more than 30,000 in September. That growth changed the bottleneck. Writing code was no longer the slowest part. Storing, comparing, testing, and merging changes became harder instead.
Why Git stopped fitting the workload
Git is a distributed system. Each clone can hold a complete repository and work offline. That design fits most projects well. Anthropic’s workload became much larger and more connected.
The company said its main-branch history was almost 100 GB. A compressed working copy exceeded 25 GB. Its continuous-integration jobs performed about 25 million checkouts each day. If those checkouts fetched data naively, they could create more than 600 petabytes of daily traffic.
Anthropic operated several hundred Git mirrors to reduce that load. The mirrors became harder to maintain as merge volume rose. They fell farther behind the latest code between refreshes. Operations involving history, such as comparing changes or finding who edited a line, also became harder on partial copies.
Most of Anthropic’s clients were agents and jobs. They were online by definition. The company decided that Git’s offline-first assumptions were working against its scale.
How Scion works
Scion puts more responsibility on an always-reachable server. The server handles history questions, differences, commits, merges, and rebases. Clients keep a bounded local cache. Many jobs can ask the server questions without keeping a complete repository.
Scion supports sparse checkouts. Those checkouts contain only requested directories. It also supports virtual checkouts. They materialize files when a job actually reads them. A changeset records changed files, while compressed file contents live in separate storage. Additional data can later support blame, path history, and search.
The design also spreads data across servers and cache layers. The company says no single server keeps a fully materialized repository copy. Immutable, content-addressed data can be cached broadly without constant coordination.
Results Anthropic reported
Anthropic reported cache-hit rates above 99 percent. It said Scion’s CPU and memory footprint was 200 times smaller than its previous Git-mirroring infrastructure. CI and research checkouts became two orders of magnitude faster, changing minutes into seconds. CI alone gained a 25 percent CPU-capacity improvement.
A seven-person team built Scion, Antique, and several related systems in five months. Anthropic also said Claude rewrote its Git-to-Scion ingestion pipeline. The rewritten version used 20 times less memory and ran four times faster. Human engineers still chose the architecture, set system constraints, and reviewed each change request. They increasingly examined the intent of a change instead of reading every line.
Why it matters
This is not a replacement for Git in ordinary projects. It is a case study in what happens when AI agents create changes at unusual speed. The slowest part of development can move from coding to infrastructure around coding.
The post also questions the usual global merge queue. Anthropic said it invested more in post-merge and pre-deployment testing because many changes can happen concurrently. That approach may fit agent-heavy development, but it can also create new review and quality problems.
The story drew 624 points and 599 comments on Hacker News. Those numbers measure community attention. They do not prove that Anthropic’s claims are correct.
What remains unknown
The measurements come from Anthropic’s own report. The post does not provide an independent benchmark or a public reproduction. Scion was designed for an enormous monorepo and many online agents. Other organizations would need to migrate code review, search, and CI around it. The post also does not show whether more than 30,000 daily merges improve software quality over the long term.
What to watch next
Anthropic expects similar pressure points to appear at other companies within six to 12 months. Watch whether Git platforms add richer server-side queries and better partial checkouts. Also watch whether human reviewers move from line-by-line inspection toward judging intent, system boundaries, and evidence from automated tests.
Anthropic built a new code manager for its AI-heavy team
📰 Full story: When AI Writes Too Much Code, Anthropic Rebuilds Source Control
Anthropic, the AI company that makes Claude, explained Scion in an [engineering article](https://www.anthropic.com/engineering/redesigning-source-control).
source control
A tool that records, compares, and restores code changes.
monorepo
One very large place holding many code projects.
Scion
Anthropic’s new system for handling code changes.
💡 The gist
- Anthropic built Scion, its own source-control system.
- It moved its huge monorepo from Git in five months.
- It says Scion saves computer power and speeds up checkouts.
Source control keeps a history of code changes. It helps teams compare work and undo mistakes. Git usually gives each copy a complete repository. That works well for many projects.
Anthropic’s project became much larger. Its main-branch history was almost 100 GB. Its compressed working copy exceeded 25 GB. Its automated testing system made about 25 million checkouts daily. A simple download method could create over 600 petabytes of daily traffic.
Anthropic ran hundreds of Git mirrors. They became harder to maintain. The mirrors could fall behind new changes. Finding old changes also became more difficult.
Most of Anthropic’s clients were agents and jobs. They stayed online. They did not need complete offline copies. So Anthropic built Scion around a server that clients can reach.
The server answers questions about history and differences. It can also make commits, merges, and rebases. Clients keep a limited local cache. Scion uses sparse checkouts. These bring only requested folders. Virtual checkouts create files when needed.
Anthropic reported strong results. Cache-hit rates exceeded 99 percent. Its new system used 200 times less CPU and memory. Checkouts became two orders of magnitude faster. They changed from minutes to seconds. CI also saved 25 percent of its CPU capacity.
Seven people built Scion and related tools in five months. Claude helped rewrite part of the system. People still chose the main rules. They also reviewed every change request.
This matters because AI agents can create code quickly. Old development tools may become the new bottleneck. But Scion is not meant for normal projects. It fits Anthropic’s enormous codebase and online agents.
The story drew 624 points and 599 comments on Hacker News. Those numbers show attention. They do not prove the story is correct.
The reported results still need outside testing. Other companies may face different problems. Watch whether Git platforms build similar server features. Also watch how humans review code in agent-heavy teams.
💬 Do AI coding teams need something beyond Git?
The comments split between people who see a real scaling problem and people who think a familiar solution may be oversold. Numbers and performance are self-reported by the article or participants, not independently measured in the thread. Source article: [Redesigning source control at Anthropic](https://www.anthropic.com/engineering/redesigning-source-control).
- Some suggest saving the AI’s instructions or a specification so the code can be recreated later. Others say the conversation and decisions along the way matter, so instructions alone are not enough.
- People debate whether AI can make the same build again, or whether any build is acceptable when it passes tests and lint. Many are cautious about calling changing outputs safe.
- Scion is often compared with Perforce, an older tool for very large teams. Some commenters think Git struggles with a huge monorepo and many AI agents; others say existing tools or a split repository might be enough.
- The article’s self-reported figures include 30,000 merges a day, 25 million CI checkouts, and a compressed working copy over 25 GB. Some commenters find this excessive, while Anthropic says large histories require caching and special infrastructure rather than repeatedly moving the whole data set.
- A self-reported result says Claude Mythos Preview made an ingestion program use 20 times less memory and run four times faster. An Anthropic participant says Git and Scion were synchronized for months, so the improvement mattered. Another user self-reports that agents may use Git history as a communication channel, although Anthropic says that is generally not allowed.
- Even with faster AI, people still choose the product, check constraints and safety, and resolve conflicts between agents. An internal success story does not by itself show that the same system will help every organization.
mature digest at 599 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
Anthropic made a new notebook for computer code
📰 Full story: When AI Writes Too Much Code, Anthropic Rebuilds Source Control
Anthropic is an AI company that makes Claude. It changed how it remembers code.
source control
A tool that remembers how code changed.
Scion
Anthropic’s new tool for handling its code.
The story
Anthropic is an AI company that makes Claude. It built Scion, a new code notebook.
Code is instructions for computers. Source control remembers code changes. Git is a popular source-control tool.
Anthropic had a huge code library. It moved that library from Git in five months.
Scion gets only the code it needs. That means less code travels around.
Claude helped write some of Scion. People chose the important rules. People checked every change request.
The company says its system became faster.
The story was discussed on Hacker News. It received 624 points. It had 599 comments. Those numbers show attention. They do not prove the story is correct.
💬 A bigger toolbox for busy AI helpers
AI helpers can make lots of code, so Anthropic built a new way to keep track of it. People still disagree about whether it is needed. The numbers and speed claims are self-reports, not independent tests. Source article: [Redesigning source control at Anthropic](https://www.anthropic.com/engineering/redesigning-source-control).
- AI instructions may not be enough to make the same software again. The whole conversation and the little decisions matter. A test can pass even when the next result is different.
- Scion is compared with an old big-team tool called Perforce. The article’s self-report says there are 30,000 code merges and 25 million automatic checks each day, so caches are needed to avoid moving huge amounts of data again and again.
- A self-reported result says AI made one program use 20 times less memory and run four times faster. A user also reports that an AI may use the code history like a message book, though Anthropic says it normally does not allow that.
- AI can write quickly, but people still choose good ideas, check safety, and fix conflicts. One company’s success story does not prove everyone needs the same toolbox.
mature digest at 599 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
💬 What HN commenters make of Anthropic’s Scion
HN commenters broadly understand why agent-heavy development might outgrow Git, but they dispute reproducibility, comparisons with existing VCSs, operational safety, and the evidence for the claimed gains. Numbers, performance claims, and behavior reports are treated as self-reports from the article authors, commenters, or Anthropic participants, not as independently verified measurements in the thread. Source article: [Redesigning source control at Anthropic](https://www.anthropic.com/engineering/redesigning-source-control).
mature digest at 599 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.