⚡ Push latency a human would never have noticed becomes the rate limiter for an agent. That's the reason GitHub is rebuilding its Git infrastructure. The numbers tell the story. Monthly Git events on GitHub doubled from 218.2 billion in September 2025 to 473.3 billion by August 2026, and September alone saw 7.38 billion commits, more than five times the year before. Pushes went from 0.69 billion to 3.35 billion, a 4.9x jump; pull request merges grew nearly 4x; GitHub Actions ran 3.26 billion times. Developers and agents working the same repository at the same time is no longer a forecast, it's in the telemetry. What surfaced underneath that growth was a structural tension in the existing design. Keeping a full copy on five fileservers by default and reaching consistency through a quorum has served a billion repositories well. But it couples durability to scale: add read replicas to absorb load and writes slow down too, because every replica participates in every push. With CI and code scanning fetching the same branch tip thousands of times a minute, that coupling becomes the ceiling. The response is to minimize coordination and split storage from compute. Only reference updates need coordinating, so object storage, validation, and secret scanning get parallelized, while compaction and garbage collection move off the serving path onto dedicated workers. Authoritative storage lives in Azure Blob Storage, with lightweight compute workers caching in front of it. Reads get absorbed by cache instead of by more durable replicas, and losing a worker degrades to a cache miss with a replacement starting immediately. Internal benchmarks report up to 35x higher write throughput. Branch protections, audit logs, and review requirements all stay intact. Title: Building Git infrastructure for agent-scale development URL: #Git# #Infrastructure#