GitHub is rebuilding its Git storage architecture from the ground up to handle unprecedented AI agent-driven demand and boost reliability.

Maintenance tasks such as compaction and garbage collection have been moved to background processes to reduce lag on the serving path.

Writes are now stored in object storage with replication managed by the storage service, while reads are directed to a dedicated channel with lightweight compute workers, coordinating primarily through reference branch pointers.

The GitHub Availability Report cites multiple incidents in spring 2026 related to performance degradation amid the surge in traffic.

The migration is designed to be non-disruptive to developer workflows, security controls, and code reviews, with no migration timeline announced yet.

The shift reflects a broader industry move toward agent-enabled development, echoing efforts at other firms refining storage and Git architectures for agent-scale usage.

Early tests of the new architecture show a 35x improvement in write throughput in internal tests, targeting growing customer demand.

GitHub’s traffic has surged due to AI agents, with monthly events rising sharply and September commits reaching about 7.38 billion, a fivefold increase from a year earlier.

The redesign decouples writes from reads, offloads maintenance, and uses object storage to handle replication and redundancy, eliminating the old three-phase commit bottleneck of the Spokes system.