What happened
Anthropic says it built and deployed Scion, an in-house source-control system designed around continuously online coding agents and jobs. The company reports that daily merged changes rose from about 2,500 in May 2026 to more than 30,000 in September. It says Scion replaced Git across its monorepo in five months and made code checkouts two orders of magnitude faster.
Anthropic frames Claude as the direct cause of a new infrastructure problem: the company says Claude can autonomously iterate on research loops, optimize programs, and build services, while existing source-control, build, and continuous-integration systems were designed for human-scale workloads. The source reports roughly 2,500 merged code changes per day in May 2026 and more than 30,000 per day by September.
The company says its monorepo’s main-branch history is nearly 100 GB, its compressed working copy exceeds 25 GB, and it performs about 25 million daily code checkouts for CI before counting developer, agent, and research workloads. Anthropic says naive Git fetching would amount to more than 600 petabytes of traffic per day, while its previous fleet of several hundred Git mirrors became increasingly difficult to operate.
Scion uses a server-centered design in which clients are assumed to be online. The server handles operations such as history queries, commits, merges, and rebases, while clients maintain bounded caches and sparse or virtualized working copies. Source-control queries are exposed through APIs, allowing some CI and automation tasks to avoid maintaining local repositories or checking out files.
Anthropic says Scion stores changesets and content-addressed file data, computes additional structures for operations such as blame and path history, and distributes data across independently scalable service tiers. It reports cache hit rates above 99%, a 200-times-smaller CPU and memory footprint than its former Git-mirroring infrastructure, checkout times reduced from minutes to seconds, and a 25% CI CPU-capacity gain. These figures are claims in the company’s source, with no outside validation provided.
The post says a seven-person team built Scion, its Antique code-review system, and related tools while completing the migration in five months. Anthropic also attributes some implementation gains to internal Claude releases, including a rewrite of an ingestion that the company says used 20 times less memory and ran four times faster. It says engineers still reviewed every change request, concentrating more on design intent and constraints than on reading every line.
Source details: anthropic.com ↗
Why it matters
The post describes a concrete infrastructure response to AI-assisted software development at unusually high volume. If Anthropic’s measurements hold, the limiting factor for agent-heavy engineering may increasingly be coordination and build infrastructure rather than code production itself. The source also illustrates that AI does not remove the need for architectural judgment: Anthropic says engineers set system constraints, reviewed intent, and redirected Claude when it produced overcomplicated or unsuitable designs. The results are Anthropic’s own reported measurements, not an independent .
This is a meaningful AI infrastructure development because the AI system is the central driver of the change, not incidental background. Anthropic is adapting a core engineering system to workflows in which agents and automated jobs generate and manipulate code at much higher frequency than conventional developer teams.
The practical implication is that organizations deploying large numbers of coding agents may need source-control systems that reduce local repository state, expose history and diff operations through APIs, and support concurrent changes without globally locking the repository. Those design choices could reduce CI overhead, but they also shift more trust and operational responsibility to centralized services.
Anthropic’s account complicates claims that coding agents make engineering expertise unnecessary. The company says Claude supplied large amounts of implementation and optimization work, while engineers selected architectural tradeoffs, enforced invariants, rejected unsuitable approaches, and managed security and compliance requirements.
The reported gains are specific to Anthropic’s unusually large monorepo and heavy internal use of Claude. The source itself warns that Scion is intended for enormous repositories with tens of thousands of engineers or large agent swarms, and that Anthropic’s willingness to spend substantial internal compute may not generalize.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
What to watch next
The key unknown is whether Scion’s architecture can benefit organizations with smaller repositories or fewer agents; Anthropic explicitly says it is not a general replacement for Git. The company does not document public availability, licensing, pricing, or a release plan. Watch for independently reproducible performance data, details about security and access-control operation, and evidence that similar systems become necessary as AI coding workloads expand.
Anthropic does not say that Scion is publicly available, nor does it provide pricing, licensing, hosted access, or a migration path for external organizations. Its suitability outside Anthropic therefore remains unknown.
Independent measurements would help establish whether the reported 200-times-smaller infrastructure footprint, two-orders-of-magnitude checkout improvement, and 25% CI capacity gain persist under different repository structures and workloads.
The system’s centralized and always-online design creates questions not answered in the source about outage recovery, offline development, portability, interoperability with Git-based tooling, and the operational consequences of its directory-level permissions and semantic review constraints.
The source describes Claude-based monitoring agents called “superfans” and automated post-merge testing, but it does not provide failure rates, staffing comparisons, or evidence that these practices are safe or effective beyond Anthropic’s internal experience.