Undo is the most important AI feature

Here is a small experiment you can run on any AI music tool. Ask it for something you are fairly sure you will not like. "Make it twice as fast and add a cowbell on every step." Then try to get back to where you were.

With a generator, you cannot go back because you never left: the fast cowbell song is a separate file, sitting next to the one you had. Fine, but you also cannot go forward, because the thing you actually wanted, your song with one change, does not exist. There are only whole songs.

With an assistant that drives a DAW, you can go back the way you go back from anything in a DAW: undo, undo, undo. An assistant turn might be thirty edits. Somewhere in the middle of them are two edits you made yourself before you asked. Good luck.

With the co-producer here, you press undo once.

Why the small feature is the big one

People talk about AI music tools in terms of the model. Which one sounds best, which one knows the most genres, which one has the nicest voice. Those matter less than they seem to, because they all decide the same thing: how good the result is on the first try.

Undo decides something more important: how bold you are willing to be.

If taking back an assistant's work costs nothing, you ask for the strange version. Darker. Half time. Turn the bassline into a countermelody. Swap every instrument. You hear five ideas in five minutes, keep one, and learn something from the four you threw away. If taking it back is expensive or impossible, you ask for small, safe things, or you stop asking and go back to doing it yourself.

The assistant with the better model and no undo gets used less than the assistant with a worse model and a perfect undo. We would bet on that every time.

Hear the difference undo makes

Here is a loop. Press play and get used to it.

Now imagine asking the co-producer for a bigger second half. This is the sort of thing it does: doubles the hats in the last two beats, adds a kick before the final snare, and lifts the last bass note.

Three changes across three lanes. In a DAW, that is at least three undo steps, probably more, tangled with whatever you did before. Here it is one step, because the turn was recorded as one unit before it was applied. Press undo and you are listening to Before again. Press redo and you are back. Toggle between them a few times and you will know, in your ears, what "a bigger second half" means. That is the lesson, and undo is what let you hear it.

What it takes to build

One-step undo of an AI turn is easy to promise and hard to keep, which is why so few tools have it. It needs three things to be true at once:

The assistant must edit through the same operations as the person. If it can reach into the project any other way, its changes cannot be grouped with yours or told apart from them. Here the co-producer has no private controls; it does what the buttons do.

A turn must be a unit before it is applied. The assistant works out everything it wants to change, the whole set is recorded as one turn, and only then does the project change. If the assistant fails halfway, nothing changed at all.

The project must be its history. Not a file that gets overwritten, but a list of everything that happened, so that any point can be returned to and any point can be forked from. Undo is then just "step back one", and a fork is "copy up to here".

None of this is exotic. It is how good software has handled edits for decades. It is only rare in AI music because most AI music tools are generators, and a generator's output has no edits to keep track of.

Undo is a feature of trust

Every other AI feature asks you to believe something about the result. Undo is the one that asks nothing. It says: try it, and if you were right to be nervous, here is the way back, and it costs one press.

We built the co-producer around that promise, and it is the reason the rest of it is safe to use. Ask for the wild version. You can always come back.