Checkpoint

A checkpoint is a saved snapshot of the working tree, and in some harnesses the conversation too, that you can return to. Before the agent starts a turn the harness records what the files looked like; if the turn goes badly, you rewind and the edits vanish. It is the undo button for agentic work, where a single turn can touch dozens of files and Ctrl+Z in the editor is no longer enough.

The need shows up the first time an agent 'fixes' something by rewriting half a module. Reviewing the diff, you realise the earlier state was better, but the earlier state is gone, spread across twenty edits. Git would have saved you, if you had committed. Most people hadn't, because they were in the middle of something. Harness checkpoints fill exactly that gap between commits. One subtlety: rewinding the files without rewinding the conversation leaves the agent believing its edits still exist, and its next tool call will act on a world that isn't there. Rewind both, or tell it what you did.

Make checkpoints boring. Commit before every substantial agent run so there is a known-good state to diff against, even if the message is just 'wip before refactor'. Use the harness's rewind for the small stuff between commits, and git stash or a worktree when you want to try two approaches side by side. A sandbox plus a clean commit is the checkpoint that never fails.