Here's a sentence that would have sounded absurd a few years ago:
"Build me an upbeat synth-pop intro, 120 BPM, with a big drop at bar 9."
Typed into an agentic music assistant, that sentence can now become a real session — separate tracks, placed clips, levels set — not a single audio file you can't touch.
The difference is the arrangement
Plenty of tools can generate a song. Far fewer can generate a song you can work on.
An agent arranges. It decides that drums enter on bar 1, the bassline on bar 5, the lead hits at the drop. Each part lives on its own track, so when you want the drop two bars later, you drag clips — you don't re-roll the whole thing.
What the agent handles
- Setup: creating the project, tempo and track layout.
- Generation: producing parts with the right model for the job.
- Placement: putting clips at the right bars instead of stacking everything at zero.
- Mixing basics: levels, panning and effects from the built-in mixer.
- Follow-ups: "make the hi-hats quieter," "add a riser before the drop," "export stems."
Where humans still win (by a lot)
Taste. Agents are fast, but they don't know that your chorus needs to feel like the last day of summer. They don't know your audience, your story or why one snare hits harder for you than another.
The best results come from a loop: you direct, the agent builds, you listen, you push back. Think of it as a very fast assistant engineer who never gets tired of your notes.
The honest limitations
- AI parts can sound generic if your direction is generic. Be specific about mood, instruments and structure.
- Agents can misread vague edits. Short, concrete requests beat long rambling ones.
- Always listen before you publish. You're still responsible for what ships.
Try it
Write one sentence describing a song. Paste it into the SoundGen assistant. Then start editing what comes back — that's where it gets fun.
SoundGen Team
Written by the team building SoundGen


