Git was built for humans — agents need an upgrade

This post introduces a next-generation source code management (SCM) system designed to address the limitations of traditional Git for agent-scale operations. It details three key problems: the 'clone tax' (excessive data transfer for single file access), 'concurrency collapse' (backend bottlenecks due to high agent load), and 'no isolation' (shared accounts and branch space leading to confusion and lack of auditability). The proposed solution involves a redesigned backend that supports the Git protocol but offers server-side querying for specific data (files, blame, history) without full clones. It features an architecture with an intelligence layer, elastic compute, and object storage, with purpose-built read/write APIs. The system is designed for both cloud and on-premises deployments and has shown significant performance gains in internal testing (up to 50x faster wall-clock time, 2x fewer tokens, 1,000x less network traffic). The post also highlights the importance of provenance and integration with the broader software lifecycle, emphasizing that a standalone Git host is insufficient.