Sebastian separates two things that often get discussed as one: a sandbox snapshot restores a machine, versioned code restores a tree you chose to record. His agent renames methods in two files and misses the third, and the way back is a push of the original worktree that lands the identical tree hash on a new commit, so the broken state stays in the log where you can read it. The measurement that carries the argument is that the whole rollback ran with no sandbox alive, at 1028 ms against 1390 ms to take a checkpoint and 2018 ms to restore one. The second half is the part to copy: fork a workspace per agent for about 170 ms, gate promotion on a deep merge preflight, and let nothing reach main that something mechanical has not approved.
We didn’t write this one — it’s Sebastian Buzdugan’s piece, published on Data Science Collective. The note above is ours; the full article is theirs.