Decoupled Multi-Agent Orchestration
Xinle Wu and Yao Lu (National University of Singapore) introduce DeOrch, a learned multi-agent orchestrator that plans without knowing which workers exist and then matches subtasks to workers, so a new worker pool can be used without retraining.
Ask this paper
Two-stage planner. A Qwen3-8B planner first decomposes the task without worker information, then picks collaboration operations using worker-identity-free matchability feedback, which allows separate credit assignment for decomposition and collaboration.
Matcher. A lightweight matcher estimates worker suitability from behavior on a fixed probe set and adapts online with a contextual bandit.
Main results. Over a six-model pool, DeOrch averages 51.5 across eight in- and out-of-distribution benchmarks, against 49.1 for MAS-Orchestra and 44.0 for Conductor, using 5.35 worker calls per task against 5.63 and 6.16.
Unseen pool. On an entirely new four-worker pool without retraining, DeOrch scores 50.2 while Conductor falls from 44.0 to 40.9, below the 47.5 of running all workers independently.
Ablations. Two-stage planning, matchability feedback and conditional credit assignment each improve results.
Abstract
Learned orchestration can automatically construct effective language-model multi-agent systems, but existing approaches couple planning to fixed worker pools and train decomposition and collaboration from the same terminal outcome, limiting transfer and obscuring credit assignment. We introduce DeOrch, which separates worker-agnostic planning from concrete worker selection. Its two-stage planner first decomposes the task without worker information, then chooses collaboration operations using compact, worker-identity-free matchability feedback from the pool, enabling conditional credit assignment to decomposition and collaboration decisions. A lightweight matcher estimates worker suitability from behavior on a fixed probe set and adapts online with a contextual bandit, allowing new workers to be incorporated without retraining the planner or matcher. Across diverse in- and out-of-distribution tasks, DeOrch outperforms prior automatic MAS orchestration methods with fewer worker calls than competing learned orchestrators, remains effective when transferred to an entirely unseen worker pool without retraining, and shows consistent gains from both components.