GitHub published a number this week that says more about what coding agents are doing to software than any survey: 7.38 billion commits in September. In August the figure was 2.9 billion. The number comes from a post by Jay Parikh, Microsoft's executive vice president for Core AI, who also oversees GitHub, and from reporting on GitHub's capacity build-out.

The infrastructure figures are just as rude. Monthly platform events rose from 218 billion in September 2025 to 473 billion a year later. GitHub has added four million CPU cores in 2026 so far, along with 160 petabytes of high-speed storage, additional GitHub Actions capacity, and 50 terabits per second of network capacity. These are not marketing numbers; they are the bill for the stampede.

The real finding, though, is architectural. GitHub's previous storage design, internally called Spokes, wrote commits through a three-phase commit and kept full repository copies on local disks. A push was therefore only as fast as the slowest replica needed for quorum. That was fine for human working rhythms for years. It is structurally wrong for thousands of agents working in parallel. In an engineering post in early October, GitHub described the rebuild: each commit is written exactly once to Azure Blob Storage, which handles replication itself; reads move onto lightweight workers; and the read and write paths coordinate over individual references rather than whole repositories. Internal tests show a 35x improvement in writes. For developers, nothing should change. GitHub gave no migration timeline.

The trigger was not flattering. In April, GitHub suffered ten performance incidents in a single month; outages and slowdowns hit Copilot too. The cause was traffic the old design never anticipated: agents write in bursts, open many branches and worktrees, commit in loops, and trigger CI on every pass. A human commits a few times a day. An agent commits hundreds of times per task.

Notably, GitHub is not alone. Cursor and a new service from a former GitHub chief executive have rebuilt their Git storage layers for the same reason. Version control — the least glamorous foundation in the industry — is being recast for machine authors.

Why it matters: 7.4 billion commits a month is not a productivity metric, it is a load profile. It says the bottleneck in agentic development is no longer the model but the substrate — storage, concurrency, rate limits, CI capacity. Treating agents as a feature is planning too small; they are a traffic pattern.

Second consequence: the cost of the default is rising. If one month of agent traffic more than doubles a platform's load, then "just host it on GitHub" stops being a neutral decision and becomes a capacity question — for GitHub, and for every company running its agents there.

The question GitHub does not answer is how many of those commits a human will ever read. The number measures activity, not progress.

Sources