Anthropic has quietly rebuilt one of the most fundamental pieces of software engineering infrastructure in use today: source control. According to a company engineering post picked up and reported by The Register on October 13, 2026, Anthropic has fully migrated its internal monorepo away from Git and onto a new, purpose-built version-control system called Scion. The company says the move was forced by a problem most engineering teams would consider a good one to have: too much code, arriving too fast, for Git to handle.
The numbers behind that problem are stark. Anthropic says it averaged around 2,500 code changes merged per day in May 2026. By September 2026, that figure had rocketed past 30,000 merges per day, a more than tenfold jump in just four months. That growth, driven largely by Claude-powered coding agents generating and submitting changes at machine speed, pushed Git past its practical limits inside Anthropic’s monorepo, according to the company’s own account of the project, published in its engineering blog and detailed at anthropic.com/engineering.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Anthropic Actually Built: Scion, Antique, and the Infrastructure Around Them
Scion is Anthropic’s name for its in-house version-control system, and it is the centerpiece of this story. Rather than patch Git to survive higher merge volumes, Anthropic chose to replace it outright for its own monorepo. The company says it began building Scion in January 2026 and completed the full migration from Git in five months, a tight timeline for swapping out the system that every engineer at the company touches every working day.
Scion was not built alone. Anthropic also built a companion code-review interface called Antique, a dedicated code search system, and a source-control identity and access management (IAM) layer. Together, those four pieces form a replacement stack for the tools most engineering orgs assemble out of Git, GitHub, and a patchwork of third-party code-review and search tools. Anthropic’s decision to build all of this in-house, rather than buy or adapt existing tooling, signals how differently the company is treating its own development pipeline now that a large share of the code flowing through it is machine-generated.
The core technical idea in Scion is the “changeset,” which functions as the atomic unit of work in the system in the same way a commit does in Git. Anthropic’s engineering team describes a changeset as an entry in a write-ahead log, and says it is designed to store most of what engineers would expect from a traditional Git commit. The company also built an API layer on top of Scion that exposes source-control queries directly, which means some automated jobs, including agent-driven ones, can interact with the repository’s history and state without needing to check out a local copy of the code at all.
Why Git Buckled Under Claude-Scale Code Generation
Git was designed in 2005 by Linus Torvalds to manage the Linux kernel, a codebase maintained by thousands of human contributors working at human speed. Two decades later, Anthropic’s internal engineering org is dealing with a different kind of contributor: autonomous and semi-autonomous coding agents built on its own Claude models, submitting changes continuously rather than in the bursts and lulls of a human workday. A tenfold jump in daily merge volume in under half a year is not the kind of load curve Git’s distributed object model, with its reliance on full-history clones and branch-merge conflict resolution, was built to absorb gracefully at a single-repository scale.
Anthropic’s own framing, as relayed in its engineering post, treats source control as something that had become the company’s biggest development bottleneck as merge volume climbed past the tenfold mark. That is a notable admission from a company whose entire commercial pitch rests on AI coding agents accelerating software development. If the tooling underneath those agents cannot keep pace with the code they produce, the bottleneck simply moves one layer down the stack, from “can the AI write the code” to “can the infrastructure absorb what the AI writes.”
This is also why the story resonates beyond Anthropic’s own walls. Any organization running AI coding tools such as Claude Code, Codex, or GitHub Copilot at scale is approaching a version of the same problem, even if their merge volumes are a fraction of Anthropic’s internal numbers. As AI-assisted commits make up a growing share of all code changes industry-wide, version-control systems built for human-paced collaboration are going to be tested in ways their original designers never anticipated.
The Scale of the Problem: Anthropic’s Merge Volume in Numbers
The headline figures from Anthropic’s disclosure are worth sitting with. Anthropic’s engineering team stated: “In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day,” a figure also independently reported by TokenPost. That is not a gradual climb. It is a step-change that lines up closely with the period in which Anthropic’s internal coding agents, including models referred to internally as Claude Mythos Preview and Claude Tag, were available for use inside the company.
| Metric | May 2026 | September 2026 | Change |
|---|---|---|---|
| Average daily code merges | ~2,500 | 30,000+ | 12x increase |
| Primary version-control system | Git | Scion (fully migrated) | Complete replacement |
| Code review tooling | Git-based review tools | Antique | New in-house tool |
| Migration duration | N/A | 5 months (started Jan. 2026) | Full monorepo switch |
| Supporting infrastructure | Standard Git tooling | Scion + Antique + code search + IAM system | 4 new internal systems |
What makes this jump particularly notable is the timeframe. Most large engineering organizations treat a version-control migration as a multi-year undertaking, often requiring parallel-running old and new systems for a transition period measured in quarters, not weeks. Anthropic’s five-month timeline, covering design, build, and full cutover of a monorepo supporting a company with billions of dollars in annualized revenue, reflects both the urgency created by the merge-volume spike and the internal priority placed on solving it before it became unmanageable.
Scion’s Design: Changesets, Write-Ahead Logs, and API-First Access
The architectural choice to center Scion on a changeset model, implemented as an entry in a write-ahead log, is a meaningful departure from Git’s content-addressable object graph. Git tracks history through a directed acyclic graph of commits, each referencing its parent, with branches as movable pointers into that graph. A write-ahead log, by contrast, is a sequential, append-only record of operations, a pattern more commonly associated with databases than with source control. That choice suggests Anthropic prioritized consistent, ordered, high-throughput writes over the flexible branching model that has made Git useful for distributed human collaboration.
Anthropic’s description of the changeset is direct: “In Scion, the unit of work is a ‘changeset,’ an entry in a write-ahead log.” The company says a changeset stores most of what would be expected from a Git commit, meaning the conceptual mapping for engineers moving from Git to Scion is reasonably close, even if the underlying machinery is different. That matters for adoption. A system that requires engineers to unlearn their entire mental model of version control is a much harder sell internally than one that mostly preserves familiar concepts while changing what happens underneath them.
The API-first design is arguably the more consequential piece for the AI-agent use case specifically. By exposing source-control queries through an API rather than requiring a full local checkout, Scion allows automated jobs, including Claude-based coding agents, to inspect repository state, history, and metadata without the overhead of cloning or syncing a working copy. For a system handling 30,000-plus merges a day, much of it agent-driven, cutting out unnecessary local checkouts for read-heavy automated tasks is a meaningful efficiency gain, both in compute and in the kind of lock contention that can plague high-frequency Git operations against a single large repository.
A Deliberately Narrow Tool, Not a Git Killer
Anthropic has been explicit that Scion is not a general-purpose Git replacement. The company’s own qualification is unambiguous: “Scion is by no means a suitable replacement for git for general use.” Anthropic has also confirmed Scion will not be open-sourced. It is a bespoke internal system, built to solve one organization’s specific scaling problem at one specific point in its growth curve, not a product aimed at the broader developer ecosystem.
That framing is a useful corrective to some of the louder online reactions to the story, which cast Scion as Anthropic preparing to challenge Git’s near-universal dominance of software version control. Nothing in Anthropic’s own disclosure supports that reading. Git remains free, open-source, and embedded in essentially every software development workflow on the planet, from solo hobbyist projects to the Linux kernel itself. Scion, by contrast, is scoped to Anthropic’s own monorepo and its own agent-heavy workload pattern, one that very few other organizations currently share at anything close to the same scale.
Still, the decision not to open-source Scion is itself informative. Anthropic has, in other contexts, chosen to keep its internal tooling tightly controlled rather than share it externally, a pattern that has drawn scrutiny before. Separate reporting, including coverage from Ground News, has noted Anthropic pursuing a DMCA takedown against a developer who reverse-engineered Claude Code, a contrast that was drawn against OpenAI’s decision to open-source its Codex CLI tool. Keeping Scion closed fits that broader posture of guarding internal engineering infrastructure, even when, as Anthropic itself says, the tool would not be especially useful to outsiders anyway.
How Anthropic’s Move Compares to the Rest of the Industry
Git has effectively had no serious rival in mainstream software development for close to a decade. Alternatives such as Mercurial, Perforce, and Subversion still have footholds in specific industries, game development and some enterprise shops among them, but Git’s combination of speed, branching model, and GitHub-fueled network effects made it the default almost everywhere else. Large technology companies with Git-scale problems of their own, including Google and Meta, have historically solved them by building custom tooling on top of Git rather than replacing it outright. Google’s internal Piper system and Meta’s Sapling project are both examples of heavily customized version-control layers that still interoperate with Git repositories at the edges.
Anthropic’s approach is more radical in one specific sense: a full migration away from Git, rather than a Git-compatible layer bolted on top. That is a materially bigger undertaking, and it is one Anthropic says it completed in five months. Whether that pays off over a longer horizon, as Scion accumulates the years of edge-case handling and tooling integrations that Git has built up over two decades, is an open question that the company’s own disclosure does not address.
| System | Approach | Built By | Public Availability |
|---|---|---|---|
| Git | Distributed object graph, branch/merge model | Open-source community (originated by Linus Torvalds) | Fully open-source, universal adoption |
| Scion | Changeset-based, write-ahead log, API-first access | Anthropic | Internal only, not open-sourced |
| Piper (Google) | Centralized monorepo with Git interoperability layers | Internal only | |
| Sapling (Meta) | Git-compatible client with custom backend | Meta | Client open-sourced, backend internal |
It is worth noting that Anthropic’s monorepo migration happened alongside a broader surge in AI-assisted coding activity across the industry. Developer surveys published this year, including the BairesDev Developer Barometer covered in our report on how developers now spend 80% of their week reviewing AI-generated code, point to the same underlying shift: AI tools are generating code faster than existing human-centric review and merge processes were built to handle. Anthropic’s Scion migration is, in effect, an infrastructure-side response to a bottleneck that most of the industry is currently experiencing on the process and review side instead.
The Claude Models Behind the Code Flood
Anthropic’s disclosure names two internal models that were available during the period covered by the merge-volume growth: Claude Mythos Preview and Claude Tag. Both are described as having been available internally during the development window, though Anthropic’s public materials do not detail their specific capabilities or release status beyond that internal availability. The naming is notable in its own right. Anthropic has used the “Mythos” label elsewhere in its public communications this year, including in the context discussed in our coverage of JPMorgan CEO Jamie Dimon’s comments on Anthropic’s Mythos AI and cyber risk, suggesting it is a model family name Anthropic has been iterating on internally across multiple contexts, not a one-off research codename.
What is clear is the directional story: as Anthropic’s own coding-capable Claude models became more available internally, the volume of code those models and their associated agents generated and submitted overwhelmed the source-control system built for an earlier, human-paced era of development at the company. That is a self-reinforcing loop worth watching. If Anthropic’s publicly available Claude models continue to improve at the pace suggested by releases such as Claude Opus 5.5, discussed in our coverage of Claude Opus 5.5 edging out GPT-6 Astra, similar merge-volume pressure could show up at any organization adopting AI coding agents aggressively, not just inside Anthropic itself.
Security and Governance Questions Raised by the New Stack
Building a source-control IAM system alongside Scion and Antique is a telling detail. Identity and access management for code repositories is not a trivial add-on; it governs who, or what, can read, write, and merge changes into a codebase that, in Anthropic’s case, is now receiving tens of thousands of agent-originated changes daily. A purpose-built IAM layer suggests Anthropic needed finer-grained control over automated agent permissions than Git’s relatively coarse branch-protection and access-token model could provide at this volume.
That governance question sits inside a wider industry conversation about securing AI-agent access to sensitive infrastructure. Security reporting this year has repeatedly flagged that the systems connecting coding agents to real repositories and real credentials are an expanding attack surface. Anthropic building a dedicated IAM layer for Scion, rather than relying on Git’s existing access controls, is consistent with that broader pattern of security teams treating agent-to-repository access as something that needs purpose-built guardrails rather than hand-me-down tooling designed for human developers.
There is also a legal dimension to how tightly Anthropic controls access to its code and tooling. A separate report from law firm Bean, Kinney & Korman, detailing what it describes as a 512,000-line Claude Code leak incident, raised questions about unauthorized access and clean-room practices around Anthropic’s coding tools, as covered at beankinney.com. Against that backdrop, Anthropic’s decision to keep Scion entirely internal, and to pair it with a dedicated IAM system, reads as part of a broader effort to tighten control over its engineering infrastructure at a moment when the company is also fending off scrutiny over how its code and tools have been accessed outside sanctioned channels.
Market and Industry Impact: What This Signals About AI-Driven Development
The Scion migration is a useful real-world data point in the ongoing debate over how much AI coding agents are actually changing software engineering, versus simply generating more code without necessarily producing more usable software. Anthropic’s own merge-volume figures, a 12x jump in four months, are a concrete illustration of AI agents dramatically increasing the raw throughput of code changes inside at least one major AI lab. Whether that throughput growth translates into proportionally more shipped product, more reliable software, or simply more code to review and maintain is a separate question Anthropic’s disclosure does not directly answer.
For competing AI labs and large engineering organizations watching this story, the practical takeaway is less about Scion itself, which Anthropic has said will stay internal, and more about the underlying signal: once AI coding agents are deployed at sufficient scale, the limiting factor on software delivery speed can shift from “how fast can the AI write code” to “how fast can the surrounding infrastructure, including source control, code review, identity management, and search, absorb what the AI writes.” That is a less glamorous bottleneck than model capability, but Anthropic’s experience suggests it becomes the binding constraint surprisingly quickly once agent-driven merge volume climbs into the tens of thousands per day.
It also raises a competitive question for companies selling AI coding tools to external customers rather than just using them internally. If Anthropic, running its own models against its own monorepo, needed a from-scratch infrastructure rebuild to handle agent-driven merge volume, customers deploying similar agents against their own codebases, even at a smaller scale, may eventually run into comparable friction with their existing Git-based workflows and review tooling.
Historical Context: From Git’s 2005 Origins to the Agent Era
Git has survived roughly two decades as the dominant version-control system by continuously adapting to larger and more demanding workloads, from the Linux kernel’s thousands of human contributors to the massive monorepos used by companies like Google and Microsoft. GitHub’s 2008 launch, and Microsoft’s subsequent 2018 acquisition of it, cemented Git’s position at the center of mainstream software collaboration for human teams. That history is precisely why Anthropic’s decision to walk away from Git for its core monorepo is notable: it marks one of the first clear, publicly disclosed instances of a major AI company concluding that Git-based workflows, battle-tested as they are, cannot keep pace with AI-agent-driven development at scale.
The timing also lines up with a broader 2026 narrative about AI coding tools maturing from assistive autocomplete into semi-autonomous agents capable of opening, reviewing, and merging their own pull requests with limited human oversight. Our earlier coverage of Codex CLI 0.162 versus Claude Code 2.1.296 tracked that same shift toward increasingly autonomous coding agents across competing platforms. Anthropic’s Scion disclosure is the clearest evidence yet that this shift has consequences reaching well beyond the coding agents themselves, down into the basic plumbing of how software gets built, tracked, and merged.
What Comes Next: Five Predictions
- Other AI labs running large internal coding-agent fleets, particularly OpenAI, Google DeepMind, and Meta, will face similar merge-volume pressure within the next 12 to 18 months and will either build comparable internal tooling or lean harder on existing custom systems like Piper and Sapling.
- Anthropic’s decision not to open-source Scion will hold, but pressure will grow from parts of the developer community for Anthropic to publish more technical detail about the changeset and write-ahead-log architecture, given the broader industry interest in agent-scale version control.
- Git itself will not be displaced for mainstream human-paced development, but expect renewed investment from the Git core team and major Git hosting platforms in performance work specifically targeting high-frequency automated commits from AI agents.
- Source-control IAM, treated as a niche concern for years, will become a more prominent enterprise security category as more organizations connect AI coding agents directly to production repositories.
- Expect more disclosures like this one from major AI labs over the next year, as internal infrastructure strain from AI-agent-driven development becomes too significant to keep entirely behind closed doors, even when, as with Scion, the resulting tools stay internal.
Frequently Asked Questions
What is Scion, Anthropic’s new source-control system?
Scion is an in-house version-control system that Anthropic built to replace Git inside its internal monorepo. It uses a “changeset” as its basic unit of work, implemented as an entry in a write-ahead log, and exposes source-control data through an API so automated jobs can query repository state without a local checkout.
Why did Anthropic replace Git with Scion?
Anthropic says its daily code merges grew from around 2,500 in May 2026 to more than 30,000 by September 2026, a more than tenfold increase driven largely by Claude-powered coding agents. The company says this growth made source control its biggest development bottleneck, prompting the move to a custom-built system.
How long did the migration from Git to Scion take?
Anthropic says it began building Scion in January 2026 and completed the full migration of its monorepo from Git to Scion in five months.
Will Scion be released publicly or open-sourced?
No. Anthropic has said Scion is a bespoke internal system built for its own needs and will not be open-sourced. The company has also stated that Scion is not suitable as a general-purpose replacement for Git.
What is Antique, and how does it relate to Scion?
Antique is the code-review interface Anthropic built alongside Scion. It was developed as part of the same broader infrastructure overhaul that also included a new code search system and a source-control identity and access management (IAM) system.
Does this mean Git is being replaced industry-wide?
No. Anthropic has been explicit that Scion is not meant to replace Git for general use. Git remains the dominant, open-source version-control system used across nearly all of the software industry. Anthropic’s move is specific to its own internal monorepo and its own AI-agent-driven merge volume.
What are Claude Mythos Preview and Claude Tag?
These are internal Anthropic model names referenced in the company’s disclosure as having been available internally during the period when Scion was developed and the monorepo migration took place. Anthropic’s public materials do not detail further specifics about these internal models beyond their internal availability during that window.
What does this signal for other companies using AI coding agents?
It suggests that organizations deploying AI coding agents at scale may eventually encounter similar strain on Git-based workflows, code review processes, and access-management systems, even if their merge volumes are far smaller than Anthropic’s internal figures.
Related
- Claude Code vs Codex vs Copilot: 65.5% Dev Share [2026]
- Dimon: Anthropic’s Mythos AI Spikes Cyber Risk 10x [2026]
- Developers Now Spend 80% of Week Reviewing AI Code [2026]
- Codex CLI 0.162 vs Claude Code 2.1.296 [2026]
- Claude Opus 5.5 Edges GPT-6 Astra by 0.9 Points [2026]
Related Coverage
- Chinese LLMs Risk $2.3T AI Capex Reset [2026]
- Anthropic Cuts AI Distillation Theft by Up to 99% [2026]
- UK Regains AI Model Access After 17-Day Freeze [2026]
- Anthropic's AI Replication Warning: 0 Rivals Respond [2026]
- Anthropic Warns AI Could Clone Itself on $55 TVs [2026]
