Long agent sessions fail in a particular way: not with an error, but with a wrong turn that is only obvious later. By the time you notice, the conversation has built on it, the files have moved on, and undoing means either arguing with the agent about what it did or unpicking the diff by hand.
A checkpoint makes that a single action. Go back to the state before the wrong turn (conversation and files together) and try a different instruction.
Conversation and files, or neither
Rolling back only the conversation leaves the agent looking at a working tree that reflects decisions it no longer remembers making, which produces the most confusing failures in the whole category. Rolling back only the files leaves it convinced of changes that are gone.
The pairing is the feature. A checkpoint that covers one and not the other is worse than none, because it looks like it worked.
Why not just git
Git is the right substrate and the wrong interface here. Committing every turn produces a history nobody wants to read and that has to be squashed before it can land; not committing means there is nothing to go back to. A checkpoint is the in-between: cheap, unnamed, local, and expected to be thrown away.
What it does not fix
Anything outside the tree: a migration already applied to a database, a request already sent, a file written outside the project. Going back in the conversation does not go back in the world, and an agent that has run a destructive command has run it.