In January I completed a course on Claude Code. In February, I moved from using agents to building around them, and created BetterAgentX, an orchestration project with twelve specialized agents.
Specialists instead of one generalist
A single assistant asked to do everything tends to do everything in the same average way. BetterAgentX explores a different approach: separate agents with defined specialties, and automatic routing that sends each request to the right one.
Memory is the hard part
Every session with a model starts close to zero unless something carries knowledge forward. Persistent memory was one of the central ideas of the project, because a team that forgets its past decisions keeps repeating its mistakes.
What building taught me
Using a tool shows you what it can do. Building on top of one shows you where it breaks. Routing can pick the wrong specialist. Memory can preserve an outdated assumption and present it as fact. More agents also mean more cost and more output for someone to review.
That is the most valuable result of the experiment. I came away with a much clearer idea of what orchestration is good for, and of where a simpler setup would be the better choice.