Agent orchestration seems to be converging toward a small set of architectural patterns.

State tracking -> tool routing -> evaluation loops

What will be interesting is how reward functions are defined once these systems operate at larger scale.