Anthropic migrated off git after merges hit 30,000 a day
The story under the story today is the infrastructure: what happens when agents write code faster than the tools built for humans, and what a lab does when the threat turns out to be inside the building.
Anthropic says its merge rate grew more than tenfold this year, so it replaced git with a version control system it wrote itself — built and fully migrated by seven engineers in five months. In May 2026, the company averaged around 2,500 code changes merged per day across its monorepo; by September it was merging over 30,000. At that volume git's central assumption — that every clone is a complete, self-contained repository — stops paying off: the history of the main branch alone is nearly 100 GB, a compressed working copy runs over 25 GB, and CI performs on the order of 25 million code checkouts per day, which fetched the naive way would mean more than 600 petabytes of git traffic daily. Anthropic had been holding that up with several hundred internal git mirrors at peak, plus partial clones, sparse checkouts, and repository snapshots, and writes that the workarounds got harder to operate as the merge rate climbed.
Its replacement, Scion, starts from the assumption that every client is an agent and therefore online. Commits, merges, and rebases happen on the server, clients keep only a bounded local cache, and there are no packfiles or repository maintenance to run. Anthropic says the new stack uses 200 times less CPU and memory than its git mirroring fleet, holds a hit rate above 99% at every cache tier, and makes CI checkouts two orders of magnitude faster — seconds rather than minutes — for a 25% capacity win in CI alone. A detail worth sitting with: an early internal snapshot of Claude Mythos Preview rewrote the company's git-to-Scion ingestion pipeline unprompted, returning a version that consumed 20 times less memory and ran four times faster while preserving the invariants the team cared about.
Do not expect to run it. Anthropic says Scion "is by no means a suitable replacement for git for general use" and would be hard to offer broadly, since adopting it means migrating the entire development lifecycle. The company makes the same argument the delivery platform Harness made in August, from the other direction — Harness rebuilds source control for the agent-written code flood.
OpenAI says it found a hidden recording device in one of its offices, used its own models to establish where it came from, and referred the matter to the FBI — which arrested the employee who planted it. In a statement Tuesday, OpenAI described the device as "sophisticated," said its models helped identify a specific employee who had placed it a few days earlier, and confirmed that person was terminated. The company says it has no indications the individual inappropriately accessed or exfiltrated model weights, source code, or customer data, and that the FBI's investigation is ongoing. OpenAI is also the only origin for the story: the suspect's name, the charges, and the model used to trace the device have not been disclosed, and no FBI or Justice Department statement has been reported — the account was independently rewritten by Reuters and Business Insider, but all three trace to the same post.
What to watch: whether the FBI confirms the arrest on the record, and how quickly other labs copy Anthropic's agent-scale tooling as their own merge rates climb.
If agents write the code and models police the office, who is actually responsible for what ships — the lab or the model? Tell us in the comments.
