GitHub, 5,000 AI agents working simultaneously… tested at 7,200 pushes per second

GitHub disclosed an internal test in which 5,000 AI agents worked concurrently in a single repository, driving 7,200 pushes per second while sustaining low-latency read requests. The company detailed the new Git storage architecture, named Rubicon, in an official engineering blog post, with design work beginning in early May 2026 shortly after an initial Azure migration. GitHub reported that average daily request volume for one repository rose roughly fourfold, from 44 million in March 2026 to 181 million in May 2026. Even the busiest repositories currently handle only tens of writes per second over short bursts, making the 7,200-push benchmark a significant stress test for AI-driven workloads. The results matter as agentic coding tools multiply concurrent commits against shared codebases, and the next thing to watch is whether Rubicon moves from internal testing into general availability.