The authors present MyAG, a graph-based framework designed to facilitate the construction and analysis of composable large language model (LLM) agent systems. The system separates agent architecture into three distinct graph abstractions: a component graph for agents and modules, a workflow graph for execution control, and a search graph for runtime execution.

  • MyAG enables flexible reuse of components with different strategies through its separation of concerns.
  • The framework supports hierarchical composition via recursive system nodes.
  • It includes monitoring and visualization tools to inspect agent execution.
  • Experiments demonstrate that the framework aids in analyzing performance-efficiency tradeoffs.

The framework is publicly available as a fully open-source project.