Cursor published a long engineering piece on 18 August by Vicent Martí arguing that hosting Git repositories at scale is a nightmare, and that the reason is in the original design rather than in any hosting provider. Linus Torvalds built Git to replace BitKeeper for kernel development, a genuinely decentralised project with many maintainers owning many subsystems, and a distributed version control system fits that workflow well.
Twenty years later the argument is that the distributed nature has become more hindrance than advantage. The average open-source project does not operate with a decentralised workflow, and the average company definitely does not: they adopt the benefits of the distributed model while running a centralised process on top of it. Which means the cost of distribution is paid by everyone and the benefit is collected by a workflow almost nobody uses.
That matters now because coding agents multiply the number of actors touching a repository, which is the shift behind local coding agents becoming shared workspaces and behind the coding-agent workbench becoming its own product layer. Repository hosting stops being neutral plumbing once agents open most of the branches. The coding agents hub tracks that pressure.
This is a vendor engineering essay published by a company that also sells code hosting, so the framing serves a product position as well as an argument. The condition to watch is whether the proposed alternatives keep a plain Git remote as an exit path, because a centralised store that improves scale while removing portability trades one nightmare for another.