HARP Framework Orchestrates Agents via Explicit Posterior Preference Inference

HARP moves agent preference beliefs from prompt text to numeric posteriors to avoid error propagation in multi-agent orchestration. This framework updates a belief over finite candidate preferences for each agent based on observed actions.