EvoMAS: Dynamic Workflow Construction for Multi-Agent LLM Systems
Researchers have introduced EvoMAS, a framework designed for constructing multi-agent workflows during execution that can adapt to changing task demands. In contrast to traditional static approaches, EvoMAS treats workflow creation as a sequential decision-making challenge along a singular task path. At each phase, it generates a clear task state through a Planner-Evaluator-Updater process and utilizes a learned Workflow Adapter to create a specific layered workflow from a predetermined set of candidate agents. This method overcomes the challenges posed by static coordination in lengthy tasks, where subgoals and information requirements evolve throughout execution. The framework utilizes large language model (LLM)-based multi-agent systems, enhancing agent specialization, tool utilization, and collaborative reasoning. Details about EvoMAS can be found in a paper on arXiv (2605.08769).
Key facts
- EvoMAS is a framework for execution-time multi-agent workflow construction.
- It addresses static coordination in long-horizon tasks.
- Workflow construction is a meta-level sequential decision problem.
- Uses a Planner-Evaluator-Updater pipeline for task state.
- Learned Workflow Adapter instantiates stage-specific workflows.
- Based on LLM multi-agent systems with specialization and tool use.
- Paper available on arXiv with ID 2605.08769.
- Overcomes limitations of one-shot workflow optimization.
Entities
Institutions
- arXiv