OpenAI and Anthropic Build Proprietary Git Backends to Handle AI Agent Traffic

OpenAI published an engineering article titled "Scaling Git at OpenAI" on November 2, revealing its self-developed Git-compatible backend, Pando. Previously, the massive increase in parallel workloads from the AI coding agent Codex caused reliability issues in OpenAI's single codebase, with fetches and pushes slowing down and merges failing, sometimes dropping reliability below 90%. Pando retains the standard Git client and protocol, replacing only the server side. According to OpenAI, median push latency fell by nearly 80% (about 5x) after implementation, with daily Git data delivery reaching nearly 1 PB and peak hourly traffic exceeding 57 TB. The system was designed by one engineer and implemented by teams of Codex agents in Rust over 10 weeks, serving production traffic since April 2026.

Anthropic announced on October 13 that it had fully migrated its internal codebase from Git to its proprietary system, Scion. Unlike OpenAI, which kept the Git client, Scion uses a distinct client-server architecture assuming users are always online, differing from Git's design for offline development. Anthropic reported that daily merge counts rose from approximately 2,500 in May to over 30,000 in September. With trunk history approaching 100 GB and CI systems performing about 25 million checkouts daily, hundreds of Git mirrors were previously required. After switching to Scion, infrastructure CPU and memory demands dropped to 1/200th of the original level, and checkout times for CI and research shortened from minutes to seconds. This system was built by a 7-person team with Claude's assistance over 5 months.

The core issue for both companies is that the speed of AI agents writing and committing code has turned traditional version control into a bottleneck. OpenAI's Pando addresses this by abandoning repository-wide consistency in favor of causal consistency, allowing unrelated branches to update independently to eliminate write contention, while storing pushed packfiles as immutable objects for reuse. Anthropic's Scion directly replaces the Git architecture, optimizing for high concurrency. Currently, both systems are internal-only and not open-source or available externally. This trend indicates that as AI agents penetrate software development, major tech companies are rebuilding underlying toolchains to handle unprecedented throughput pressures.