When AI Moves the Bottleneck

On October 13, Anthropic published an interesting engineering story about a problem created by its own success with AI coding.

In May 2026, Anthropic was merging around 2,500 code changes a day. By September, that number had crossed 30,000.

Claude had made writing code so much faster that coding itself was no longer the bottleneck. The problem had shifted to what happened after the code was written.

Their existing infrastructure, including Git, builds and continuous integration, was struggling to keep up. Git is designed around assumptions that work well for human developers, including maintaining local copies of repositories and supporting offline work. At Anthropic’s scale, with AI agents continuously generating code, those assumptions were becoming expensive.

They had already tried workarounds, including hundreds of internal Git mirrors. But these were becoming increasingly difficult to manage.

So Anthropic built Scion, its own source-control system, designed for this new scale of AI-assisted development. A team of just seven engineers, working extensively with Claude, built Scion and a new code-review tool and completed the migration in five months.

There is an interesting irony here. Claude helped create the problem, and Claude helped build the solution. But the bigger learning goes well beyond software engineering.

When any form of automation dramatically accelerates one part of a workflow, the bottleneck doesn’t necessarily disappear. It often moves somewhere else.

Some examples:

• You can generate reports faster, but approvals may still take days.
• You can analyze data faster, but decisions may still wait for the next review meeting.
• You can write code faster, but the systems that test and deploy it may not keep pace.

Making one task faster doesn’t automatically make the entire workflow faster.

And perhaps this is where organizations need to pay more attention. Not just to what AI can accelerate, but to what needs to change around it.

Reference: Anthropic — Redesigning Source Control