A multi-agent prompting system coordinates several models or roles to produce, critique, filter, vote on, or synthesize an answer. Instead of trusting one generation, it distributes parts of the task across cooperating agents that can cross-check one another. section 7.6
Generate, critique, synthesize
One model generates options, a second critiques or filters them, and a third votes or synthesizes the final output. Constitutional AI and collaborative debate models illustrate how this arrangement can improve factual consistency, alignment, and robustness, while not eliminating uncertainty. section 7.6
The technique extends prompt engineering from writing a single input to orchestrating a full cycle of construction, iteration, optimization, and structured interaction. Its benefits depend on how well the roles and constraints are designed. section 7.6
