The point about labels being a snapshot is the one I'd stress for the later classifier stage. When you do start collecting labels, keep the capability catalog version next to each one and hold out a test set per version, so a drop in accuracy after onboarding a new agent shows up as a routing regression and not as noise. Class imbalance will also bite early: a forecasting agent that gets 2% of traffic can look fine on overall accuracy while failing most of its requests, so report per-route recall alongside the abstain rate. A small hand-written set of the ambiguous multi-domain questions, like your sales-are-down example, makes a useful fixed regression suite.