Addressing AI Agent Commit Bottlenecks: OpenAI and Anthropic Develop Custom Version Control Backends
AI assistants · AI coding
To address performance bottlenecks from high-frequency code commits by AI agents, OpenAI built the Git-compatible backend Pando, while Anthropic created the client-server system Scion. Both were built within months using internal models, drastically improving throughput and latency, though neither is publicly available or open-source.
As AI agents increasingly write and commit code, traditional version control systems face severe performance bottlenecks. Engineering posts from OpenAI and Anthropic detail their respective internal solutions. OpenAI built Pando, a Git-compatible server backend that preserves existing Git clients and protocols, reducing median push latency by nearly 80% with a cache hit rate exceeding 99%; developed within 10 weeks with Codex assistance, it has handled production traffic since April 2026. Anthropic replaced Git with Scion, a client-server architecture built for always-online setups, handling merge surges from roughly 2,500 daily in May to over 30,000 in September while cutting CI checkout times to seconds and reducing CPU/memory footprint to 1/200. Both systems remain strictly internal and are not offered publicly or open-sourced.
Key takeaways
- OpenAI deployed Pando, a Git-compatible backend that retains existing clients while cutting push latency by nearly 80% and handling over 57 TB per peak hour.
- Anthropic replaced Git with Scion to manage over 30,000 daily merges and 25 million checkouts, cutting infrastructure CPU and memory demands to 1/200th of previous levels.
- Both systems were built within months with internal AI assistance (Codex and Claude), but both remain internal-only with no plans from Anthropic to open-source Scion.
Source：INSIDE Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.