An analysis of AI development infrastructure separates bottlenecks across Git clients, protocols, storage and collaboration. OpenAI’s Pando retains Git compatibility while handling branches independently and reusing pack data, whereas Scion uses a dedicated client and server-centered operations. OpenAI’s own account also describes a trade-off: reading several branches can expose states from different moments. Accepting many agents’ pushes is distinct from validating and integrating changes into main. Conversation records kept separately may also be absent from a simple code clone.