AI agents now write code faster than the systems built to store it can handle. Anthropic and OpenAI both hit that limit this year. Both rebuilt their source control. One kept Git. One replaced it.
The numbers
Anthropic merged about 2,500 changes a day in May 2026 and more than 30,000 a day in September 2026 (Anthropic Source Control Team, "Redesigning source control at Anthropic," Anthropic Engineering, October 13, 2026). Its main branch history is close to 100 gigabytes. Its testing systems check out code about 25 million times a day.
OpenAI's commits to its main branch grew 1,378 percent from the first quarter of 2024 to the third quarter of 2026, with 2,442 percent expected by the end of 2026 (Aaron Friel, Taylor Blau and Ted Nyman, "Scaling Git at OpenAI," OpenAI, November 2, 2026).
The result was the same at both companies. Anthropic says source control became its biggest development bottleneck. OpenAI says its Git backend fell below 90 percent reliability, with slow fetches and failed merges.
Agents caused this. Claude and Codex write and submit code in parallel, continuously. Writing code is no longer the slow part. Storing, merging and testing it is.
Why Git breaks here
Git was built in 2005 for human developers working apart, often offline. Every user gets a full copy of the repository, including its full history. The server builds a custom packfile for each request. Changes wait in a single order to keep the repository consistent.
Each of those choices costs little at human scale and a lot at agent scale. Full copies mean Anthropic had to run several hundred Git mirrors just to serve requests. Custom packfiles mean the same work is repeated millions of times. A single order means parallel work waits in one line.
Agents do not need Git's main benefit. They are always online and work inside data centers. They do not need a full copy of a repository on a local machine.
OpenAI kept Git and rebuilt the server
OpenAI's system is Pando, a Git compatible backend in production since April 2026. Developers and agents still use standard Git commands. No one had to learn a new tool.
Pando changes two things. It coordinates each branch separately, so unrelated branches move in parallel and only pushes to the same branch wait. And it stores packfiles instead of building a new one for each client, then serves them from a regional cache called Express, with hit rates above 99 percent.
OpenAI reports median push latency down nearly 80 percent, merge reliability at 99.99 percent, nearly a petabyte of Git data delivered per day, and more than 57 terabytes in its peak hour (Friel, Blau and Nyman, "Scaling Git at OpenAI," OpenAI, November 2, 2026). One engineer designed the core. Teams of Codex agents implemented it in ten weeks.
Anthropic replaced Git
Anthropic's system is Scion, built by seven people in five months. The whole company migrated off Git. The same team also built a code review tool, Antique, plus code search and access control, because replacing Git meant replacing the tools built on top of it.
Scion reverses Git's design. The client is assumed to be always online. The server performs commits, merges, rebases and history queries. Clients keep only a small local cache.
Anthropic reports Scion uses 200 times less processing capacity and memory than its old Git mirrors, checkouts are about 100 times faster, and its testing systems recovered 25 percent of their capacity (Anthropic Source Control Team, "Redesigning source control at Anthropic," Anthropic Engineering, October 13, 2026). Claude helped build it. An early Claude Mythos Preview rewrote the Git to Scion ingestion pipeline to use 20 times less memory and run four times faster.
What this means
Most teams do not need either system. Anthropic itself says Scion is not a suitable replacement for Git for general use. It only makes sense for enormous monorepos worked on mainly by agents.
Three things follow for everyone else.
First, the bottleneck in software development has moved from writing code to verifying it. Review, build, testing and security now face the same load source control just did.
Second, both systems are private. If agent scale infrastructure stays inside the labs, the labs' advantage compounds, because their agents also built the infrastructure. Anthropic says it wants to bring Scion's ideas back to Git with GitHub, GitLab and the open source community. OpenAI says it is exploring sharing Pando's technology and has submitted fixes to Git itself.
Third, Git is not obsolete. It is being outgrown at the top end, by the two companies pushing agent coding furthest. Everyone else will feel this later, in weaker form, as agent coding spreads.
Git lasted 20 years because it matched how humans write code. Anthropic and OpenAI are now rebuilding source control to match how agents write code. The open question is whether those rebuilds stay private or become the next version of the shared tool.
Sources
Aaron Friel, Taylor Blau and Ted Nyman, "Scaling Git at OpenAI," OpenAI, November 2, 2026.
Anthropic Source Control Team, "Redesigning source control at Anthropic," Anthropic Engineering, October 13, 2026.