EvoMAS: Heuristics in the Loop—Evolving Smarter Agentic Workflows
Abstract
The rapid development of Large Language Models has driven Multi-Agent Systems (MAS) growth, but constructing efficient MAS still requires labor-intensive manual design. Current automation methods often generate templated agents, rely on monolithic optimization, and ignore task complexity gradients. This paper presents Evolutionary MAS (EvoMAS), a biologically inspired framework that addresses these limitations through three interconnected dimensions: (1) dynamic and diverse evolutionary strategies with six biologically inspired operators (3 exploration, 3 exploitation) and adaptive strategy selection; (2) role-level evolution that dynamically optimizes agent specialization and collaboration patterns; and (3) curriculum-guided evolution that partitions tasks by difficulty and evolves sequentially from simple to complex under cross-stage stability constraints. To bridge the inefficiency of pure evolutionary search and the rigidity of manual design, we introduce the Cyber Creator, a meta-control system that combines dynamic rule formulation with reflective updates. Experiments show EvoMAS consistently outperforms existing methods across multiple domains while remaining cost-efficient, with agent roles evolving from homogeneous actors to specialized reasoning ensembles.