Anthropic's AI code flood outgrows Git⁠.

Anthropic has published an account of how it replaced Git, the tool software teams use to track every change to their code, with an in-house system called Scion.¹ The company says that in May 2026 it averaged around 2,500 code changes merged a day, and that by September it was merging over 30,000 a day.¹ It says coding is no longer what holds it back, because Claude can build and improve software by itself, and that the tools built for human scale, source control, builds and testing, are now the bottleneck.¹

The numbers behind that are large. The history of its main code base is almost 100 GB, a compressed working copy is over 25 GB, and it runs on the order of 25 million code checkouts a day for automated testing alone.¹ At its peak it ran several hundred internal copies of the repository to cope with the load.¹ Because nearly all of its users were agents and automated jobs that are always online, it built a system that keeps the work on a central server and lets each client fetch only what it needs.¹

A team of seven people built Scion and a code review tool called Antique, and moved the company across in five months, while the amount of code merged each day rose tenfold.¹ Anthropic says the new set-up needs 200 times less computing power and memory than the old one, and that checkouts now take seconds rather than minutes.¹

The company says every engineer on the project now rates Claude as better at writing code than they are, but that very often the engineers' job was to tell Claude no.¹ It says agents sometimes broke core rules of the system, and that one automated review led Claude to add needless checks and complexity.¹ People still review every change, but they now look at the intent of a change rather than reading every line.¹

Anthropic says Scion is no replacement for Git for general use, and makes sense only for enormous code bases with tens of thousands of engineers or swarms of agents.¹ It also says its own spending on AI for internal use far exceeds what any other business would realistically want to spend.¹ Simon Willison's weblog describes Scion as sounding like a reinvention of centralised source control, along the lines of Subversion, and guesses that Anthropic's newest models are a key factor in the faster pace.²

Your tools were built for people

This is a story about Anthropic's own plumbing, and you will not be running Scion. Anthropic says as much, and admits it spends far more on AI for its own use than a normal business would. There is no British angle here. The useful part is the pattern underneath. When AI does more of the work, the hold-up moves to whatever was built for human pace: the approval step, the review, the system of record. A firm that lets agents raise invoices or update customer records may find the same thing on a smaller scale, and sooner than it expects.

The lesson we would underline is about review. Anthropic's engineers did not stop checking the work. They changed what they check, from each line to the intent, and much of their job was saying no. That is the right shape for any business putting agents to work. Decide the rules an agent may not break, check that it keeps them, and keep a person on the changes that matter.

Also today

• Nvidia gives the UK $250m of computing for AI safety

  Nvidia is donating $250m of computing power over five years to support AI safety research in the UK, delivered through UK datacentres, with a share going to university labs and Imperial College London.³

• Half of firms report an AI agent going off script

  In a Salt Labs survey of more than 380 security and technology professionals, half said an AI agent in their organisation had taken an action in the past year that was not intended or authorised.⁴

• FBI shuts 40 laptop farms in a North Korean jobs scam

  The FBI has shut down 40 laptop farms used by North Korean fake IT workers who won remote jobs with stolen or fabricated identities, seizing 187 devices and arresting two people.⁵

• Scimplify raises $25m to make specialty chemicals with AI

  Indian start-up Scimplify has raised $25 million, in a round led by Hitachi Ventures, to expand its AI-powered approach to producing specialty chemicals and critical materials as manufacturers try to rely less on Chinese suppliers.⁶

• AI model stripped of its safeguards within a day

  A report from the UK nonprofit Tech Against Terrorism found that within 24 hours of Alibaba releasing a small open-weight model, versions with their guardrails removed had appeared online.⁷