Learned orchestration can automatically build effective language‑model multi‑agent systems, yet current approaches tightly couple planning to a fixed worker pool and train decomposition and collaboration from the same terminal outcome, which hampers transferability and obscures credit assignment. We introduce DeOrch, which decouples worker‑agnostic planning from concrete worker selection. Its two‑stage planner first decomposes the task without any worker information, then selects collaboration operations using compact, worker‑identity‑free matchability feedback from the pool, enabling conditional credit assignment to both decomposition and collaboration decisions. A lightweight matcher estimates worker suitability from behavior on a fixed probe set and adapts online via a contextual bandit, allowing new workers to be added without retraining the planner or matcher. Across diverse in‑distribution and out‑of‑distribution tasks, DeOrch outperforms prior automatic MAS orchestration methods with fewer worker calls, transfers effectively to entirely unseen worker pools without retraining, and shows consistent gains from both components.
Review