EvoMAS: Learning Execution-Time Workflows for Multi-Agent Systems

Researchers introduce EvoMAS, a framework that dynamically constructs multi-agent workflows during task execution rather than using static, pre-optimized designs. The system uses a Planner-Evaluator-Updater pipeline to assess task state and adapts agent coordination across execution stages, demonstrating superior performance on complex reasoning tasks compared to existing approaches.

EvoMAS addresses a fundamental limitation in current LLM-based multi-agent systems: the assumption that a single workflow design can handle diverse, evolving task requirements. Traditional approaches optimize agent coordination upfront and apply it uniformly, which fails when tasks involve changing subgoals and emerging information needs. This research shifts the paradigm toward adaptive, execution-time workflow construction.

The framework's innovation lies in treating workflow design as a sequential decision problem tied to actual task progression. By maintaining explicit task states and training a Workflow Adapter with policy gradients, EvoMAS learns when and how to reconfigure agent teams dynamically. The use of sparse terminal rewards as primary supervision, with supplementary process rewards analyzed separately, reflects practical constraints in evaluating multi-step agent reasoning.

For the AI industry, this work signals growing sophistication in multi-agent system design. As LLM applications tackle increasingly complex domains—research, decision-making, problem-solving—static coordination becomes a bottleneck. EvoMAS's demonstrated improvements on benchmarks like GAIA and DeepResearcher validate that adaptive approaches yield tangible gains. The separation of concerns between task-state construction and learned adaptation suggests these components provide independent value, enabling future modular improvements.

Looking ahead, the critical question is whether execution-time adaptation scales to longer horizons and more diverse task distributions. The framework's computational overhead during execution, training requirements for new task domains, and generalization across agent pools remain open questions. Industry adoption likely depends on whether EvoMAS can operate efficiently within production cost constraints while maintaining interpretability of adaptive decisions.

• →EvoMAS enables multi-agent workflows to dynamically adapt during task execution rather than following static pre-optimized designs.
• →The framework outperforms single-agent baselines and existing automated multi-agent workflow methods on complex reasoning benchmarks.
• →Explicit task-state construction and learned workflow adaptation provide complementary benefits for handling evolving task requirements.
• →Process rewards become particularly valuable in extremely sparse-reward settings where terminal task success is rare.
• →The approach treats workflow configuration as a sequential decision problem optimized via policy gradients with verifiable task success signals.