I want to work with agents the same way I work with people: through an ongoing relationship. I explain what I'm trying to do, we work through the idea, and eventually I ask for an outcome. I don't want to pick a model from a menu before I'm allowed to ask a question. For each project, I want that relationship to remain stable: it should know what we're building, what we've decided, and what's already moving.
The chief of staff does the organizing behind that relationship. Small requests can stay in the conversation. Substantial work moves into a focused thread with its goal, constraints, relevant context, authority, and the decisions that still need me. The work gets room to develop without turning the main conversation into a junk drawer.
I've been building Maverick around this premise. Its context directory keeps the project's plans, decisions, and artifacts together. The chief of staff can answer directly, delegate to a specialist, or start a fresh run of itself with a narrower context. The model can be the same; the separation comes from giving the work its own thread. When the work is done, the result comes back to the project conversation.
The pattern is recursive because the interface stays the same. Once one chief of staff can accept a request, organize the work, and return a result, another can wrap it. A project chief can manage task threads. Another chief can manage several project chiefs.
That leaves me in one conversation about the company while the chief of staff organizes the rest. I've spent enough time being the router.