GitHub is Rebuilding Its Git Infrastructure for Agent-Scale Development

GitHub announced that it is rebuilding its core Git infrastructure to support agentic software development and high-volume workloads. According to the company, total Git activity on the platform more than doubled between September 2025 and August 2026, reaching 473.3 billion events per month, while monthly commits surged over five times to 7.38 billion in September.

Architectural Challenges at Scale

Agentic workflows introduce distinct scaling hurdles:

• Commit Turnaround: AI agents operating in tight loops commit or checkpoint frequently, making push latency a primary bottleneck.
• Write Throughput: Monthly pushes grew 4.9x year-over-year to 3.35 billion, concentrating heavy write activity onto single repositories.
• Merge Contention: Trunk-based development and pull requests funnel extensive activity onto single references.
• Read Fan-Out: CI pipelines and code scanning clone or fetch branch tips thousands of times per minute, supported by 3.26 billion GitHub Actions runs in September.

Current Architecture Limitations

GitHub currently stores repositories using Spokes, maintaining full copies on multiple local fileservers. While this provides low-latency native reads and redundancy via a three-phase commit quorum, it couples durability with scale. Adding read replicas increases write overhead because every replica participates in every write.

Redesigning for the Agent Era

To decouple scale from durability, the new infrastructure follows core distributed systems principles:

• Coordinate Only Necessary Updates: Reference updates require agreement, but underlying object storage and validation can happen in parallel.
• Move Maintenance Off the Serving Path: Heavy tasks like garbage collection and compaction will run on separate workers against durable storage.
• Independent Scaling: Authoritative repository data will live in Azure Blob Storage, while lightweight compute workers handle read traffic and cache data.
• Faster Failure Recovery: Losing compute workers functions similarly to a cache miss, allowing rapid replacement without rebuilding full repository copies.

Internal benchmarks indicate the new architecture has achieved up to 35 times higher write throughput. GitHub noted that it is rolling out these foundational changes without altering standard developer workflows, compliance controls, or security governance. Further technical details on the architecture will be shared in upcoming posts.