Why we gave each agent one job and refused to add more

The first version of Symphony was a single agent that tried to do everything. Campaign strategy, royalty audits, content briefs, sync pitching. We thought the leverage was in breadth. One AI, infinite surface area.

It was a disaster. Not because the model was bad. Because nine jobs in one prompt makes a generalist with good vibes and mediocre output. It sounds smart. It rarely is.

The fix: scope each agent to a single function and give it the tools, memory, and context to be genuinely expert at that one thing. Maude only reads royalty statements. She does not write tweets, plan tours, or give career advice. She finds money you are owed. That is it.

“A scoped agent that does one thing well is worth ten assistants that do everything adequately.”

The discipline changed how we build. One scope means a precise memory file, predictable tool calls, diagnosable errors. You know exactly what an agent should and should not do, so you know immediately when something is wrong, and you can fix it without tearing the system down.

It also changed what clients trust. “Maude is our royalty person” lands differently than “our AI does royalties.” One is an accountability structure. The other is a vague promise. In music, trust is the entire game.

Nine agents, nine jobs. The composability lives in the memory layer: agents that share context through a master file, not one model context-switching across an entire career.