Anthropic details Scion, a git replacement for its monorepo, built by seven engineers with Claude in five months, as daily merges rose from 2.5K in May to 30K+
As our merges per day grew more than tenfold, source control became our biggest development bottleneck.
Anthropic
Context & Ripple Effects
Anthropic had already been increasing the parallelism of Claude-assisted engineering: dynamic Claude Code workflows were designed to run hundreds of subagents, while the company said Claude authored more than 80% of merged code in its codebase. Its earlier use of 16 parallel agents to produce a 100,000-line Rust C compiler demonstrated the same shift from model assistance to coordinated software production.
Scion is Anthropic’s infrastructure response to that throughput change. With daily merges rising from roughly 2,500 in May to more than 30,000 in September, the company says source control—not code generation—became the limiting development system.
First-order effects
Anthropic replaces Git for its monorepo with Scion, giving its engineering organization a source-control system built around its higher merge volume and Claude-assisted workflow.
The seven-person Scion team demonstrates that Claude can compress the staffing and elapsed time needed to build internal developer infrastructure, not only application code.
Second-order effects
Anthropic’s bottleneck shifts from generating and merging code toward operating review, policy, and coordination processes that can safely handle agent-driven change volume.
Established Git and code-review workflows become a weaker fit for Anthropic’s internal environment, raising the value of tooling designed for high-frequency, agent-mediated merges.
Third-order effects
If other large AI labs reach comparable code-change rates, source control and review may evolve from developer utilities into an agentic development control plane that governs how automated work enters production.
The competitive advantage in AI-assisted engineering may increasingly depend on proprietary workflow infrastructure as well as model capability, because legacy collaboration systems can become throughput constraints.
The trend: Agentic coding is pushing software organizations to redesign the developer control plane around coordinating and validating machine-generated changes at scale.
“Very often, we found that the role of the engineers was to tell Claude ‘no.’” This is now the job! Push the model out of distribution when describing what you want and when reviewing its work! Our work is communicating and enforcing our intent.
Talking about recursive self improvement: “Claude is able to autonomously iterate on research loops, micro-optimize complex programs, and build new services from the ground up at a speed and scale we haven't seen before.” Even if we haven't quite reached RSI yet, it's incredibly …
New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, w…
Anthropic's code changes merged per day has 10x since May. 'Claude is able to autonomously iterate on research loops, micro-optimize complex programs, and build new services from the ground up at a speed and scale we haven't seen before.'
One year ago, I would have thought replacing an entrenched source-control vendor with homegrown source control and code review systems was a terrible idea. I saw these projects take years before. It was done in 5 months and is supporting 10X the volume. https://www.anthropic.com/…
“In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day.” 12x growth in just 4 months. This explains why GitHub keeps breaking.