MASCraft: Multi-Agent Specialization through Role–Skill Composition and Self-Evolving Orchestration

Effective multi-agent systems (MAS) require agents specialized for different types of work. However, role prompting primarily specifies who an agent should act as, without necessarily providing sufficient procedural guidance on how to perform its work. We introduce MASCraft, a framework that combines role–skill composition with self-evolving orchestration for multi-agent specialization. By pairing independently defined roles with reusable skills, MASCraft equips specialists with both professional identities and explicit procedural guidance. An orchestrator agent plans subtasks, selects role–skill pairs from separate libraries, and organizes collaboration according to information dependencies. To better meet the specialized demands of MAS orchestration, we provide an explicit policy that evolves through execution experience. This policy comprises two textual modules: Environment Knowledge, which records task characteristics and recurring failures, and Team Strategy, which guides specialist configuration and coordination. MASCraft generates multiple rollouts per task, selects tasks with contrasting results for within-task reflection, and aggregates lessons across tasks to update the policy, guiding subsequent team construction. Extensive experiments on HLE, DeepResearch Bench, and BrowseComp-Plus show that MASCraft consistently outperforms competing MAS methods across different LLM backbones. Code is included in the supplementary material.