copy-on-write

Copy-on-write makes something look like it's been copied, while actually delaying the real copying until it's truly needed. It's like handing out a single contract to several people as if each had their own copy—but really, everyone is looking at the same original document. Only when someone actually tries to edit even a single line does a genuine copy get made just for them. If no one touches it, no copy is ever created.

This approach is useful because of speed and space. Copying an entire file or program's full state every single time takes a lot of time and storage. But if only the changed parts get saved later, the cost of copying drops dramatically. So when you need to roll back to an earlier point in a process, you don't have to have made a full duplicate of that state in advance—you can rewind lightly, only when it's actually needed.

The article describes an AI agent execution recording tool called Shepherd, noting that it "saves the agent process and filesystem state together using copy-on-write." This is said to make rolling an agent back to a specific point in time much faster and cheaper. It's worth noting, though, that copy-on-write isn't a technology unique to Shepherd—it's a general storage optimization technique that's been used across computer systems for a long time.