GitHub Rebuilds Git Infrastructure to Support Agent-Scale Development

GitHub is rebuilding its Git infrastructure to handle the surge in agentic software development, where agents and developers work concurrently in repositories. Agentic development has driven total Git activity to 473.3 billion events per month by August 2026, more than double the previous year's level. In September alone, developers and agents made 7.38 billion commits, over five times the volume from a year earlier.

Architectural Bottlenecks

The current architecture, which uses Spokes to keep full repository copies on local disks, faces critical limits at high scale. Because durability and scale are coupled, adding read replicas slows down writes, as every push must be acknowledged by a quorum of replicas. Key challenges include:

• Commit turnaround: Agents bound by push latency create bottlenecks.
• Write throughput: Pushes grew 4.9x year over year, converging on single points in the architecture.
• Merge contention: Trunk-based development funnels work onto single references.
• Read fan-out: CI and code scanning multiply each push into thousands of reads, with GitHub Actions running 3.26 billion times in September.

New Design Principles

The new architecture separates durability from scale to allow independent scaling of reads and writes. Authoritative repository data will move to Azure Blob Storage, while lightweight compute workers handle requests and caching. This design minimizes coordination by limiting consensus to reference updates, allowing object storage and validation to happen in parallel. Maintenance tasks like compaction and garbage collection are moved off the serving path to background workers.

Performance and Impact

Internal benchmarks indicate the new system delivers up to 35 times higher write throughput, with read capacity scaling independently to meet demand. The rebuild aims to preserve existing controls like branch protections and audit logs while supporting the most demanding workloads, such as enterprise teams running thousands of agents against a single codebase. This foundation is intended to benefit all users, from large enterprises to individual open-source maintainers, by providing a faster, more resilient platform.