In Intelligence is eating software I started from one idea: when writing code costs little, value and problems move elsewhere. Yesterday, 13 October, Anthropic published on the blog of its Engineering group a concrete case of where they move. Between May and September the changes merged every day into its monorepo went from about 2,500 to more than 30,000. According to Anthropic the bottleneck has become the infrastructure built for a typically human speed: source control, builds and continuous integration.

When git becomes the bottleneck

The figures describe the scale. The main branch history weighs almost 100 GB, a compressed working copy exceeds 25 GB, and continuous integration alone performs about 25 million checkouts a day, not counting developer environments, agent sandboxes and research jobs. Fetching the repository naively, Anthropic estimates, would mean more than 600 petabytes of git traffic a day.

Git was born decentralised: every clone is a complete repository that also works offline. Anthropic used it instead as a centralised system, with almost all clients always online, and kept things going with partial clones, sparse checkouts and, at peak, several hundred internal mirrors. As changes grew, that architecture had become hard to run.

Version control built for agents

Anthropic’s answer is Scion, an in-house version control system built on six principles.

  • Clients are always online. The client fetches what it needs when it needs it, instead of holding a complete copy.
  • Work happens on the server. History, blame, commits, merges and rebases run on the server; the client keeps a bounded cache and lays out the working files.
  • Source control as a service. Every query goes through an API, and many jobs, such as the automated bisect and revert pipelines, need no local repository.
  • Horizontal scalability. Data is distributed, no complete copy exists anywhere, not even on the server, and there are no packfiles and no maintenance.
  • Cheap, lock-free commits. The unit of work is a changeset, an entry in a log written on the server, which lists the files changed relative to the parent and points to content stored by hash. Updating a ref is an atomic operation on a single Spanner row, and branches, apart from main and release branches, are discouraged.
  • An optional, sparse working copy. CI can ask the server for diffs directly, and checkouts materialise only the files that are read.

Traffic flows through three tiers that answer the same calls: a daemon per node, a cache per region and the servers. The separation is strict, so the failure of a regional cache does not overwhelm the servers upstream. Anthropic cites Google’s Piper and Meta’s Sapling as prior art.

The results

Seven people built Scion and migrated the whole monorepo in five months. In the same period they also built a code review system called Antique, code search and a per-directory permission system.

  • An infrastructure about 200 times smaller than the old mirrors, in CPU and memory.
  • CI and research checkouts about a hundred times faster, from minutes to seconds.
  • 25% of CPU capacity recovered in CI alone.
  • A cache hit rate above 99% at every tier.

An early internal version of Claude rewrote the pipeline that imported data from git: the new one uses 20 times less memory and runs 4 times faster, with the same ordering constraints.

Anthropic is explicit about the limits of the result. Scion is not a general replacement for git, it only makes sense for very large monorepos with many engineers or agents, and adopting it means migrating the whole development lifecycle. The company will not release it: it wants to share the ideas with GitHub, GitLab and the open-source community, and expects problems at this scale to become common within six to twelve months.

The engineers’ work changes

The most interesting part of Anthropic’s account is about people. Engineers bring the understanding of the system and the design choices, and often their job was to tell Claude «no», for example when it proposed brute-force scans of the commit log instead of querying the review system. Today they review the intent of changes, not every line.

Day-to-day operations follow the same logic. Dedicated Claude instances watch individual components by reading telemetry and logs, and according to Anthropic they catch problems that static alerts miss. The merge queue with a global lock was dropped because it does not hold at agent scale: changes are merged and then tested before deployment. Concurrent changes can conflict, but Claude usually resolves the conflicts before a person notices them.

What I think

The two-speed era is beginning, and software is the first sector where it shows. There is an agentic lane, AI today and superintelligence in prospect, that produces, merges, verifies and monitors at a pace of tens of thousands of changes a day. And there is a human lane, which decides intent, reviews design choices and answers for the results. Anthropic’s case shows that both lanes already exist, in the same company and on the same code.

This is why governing the agentic lane has to become the absolute priority of the human one. Practices born for people writing and reviewing code, line-by-line review, the merge queue, manual maintenance, do not keep pace, and at Anthropic conflicts are already resolved before anyone sees them. Staying one step ahead, as we have always tried to do with security and controls, may no longer be possible.

Trying requires new paradigms, which do not have to start from scratch. The tools exist and there are many of them: the EU AI Act, the NIST AI RMF, ISO/IEC 42001, the OWASP lists for agentic applications, the risk management models of every sector. The problem is holding them together and making them verifiable at the speed of the first lane. That is why I proposed OISG, Open, Intelligent, Secure, Governed: a vocabulary that indexes these frameworks and turns their obligations into measurable properties, with evidence that can be produced automatically.

Cover image: The Horse in Motion by Eadweard Muybridge, 1878. The mare Sallie Gardner, owned by Leland Stanford, galloping on 19 June 1878 on the track of his Palo Alto farm, where Stanford University would be founded a few years later. Muybridge photographed her with twelve automatically triggered cameras, capturing for the first time a movement too fast for the human eye. In the last frame the horse is standing still — public domain — https://commons.wikimedia.org/wiki/File:The_Horse_in_Motion_high_res.jpg