New 2D Framework Classifies LLM Agent Design Patterns

A new two-dimensional classification framework for LLM-based agent architectures has been introduced by researchers. This framework features a Cognitive Function axis, which includes seven categories: Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, and Governance, alongside an Execution Topology axis comprising six archetypes: Chain, Route, Parallel, Orchestrate, Loop, and Hierarchy. The resulting matrix of 7x6 reveals 27 distinct patterns, with 13 being newly named. Current frameworks from Anthropic, Google, and LangChain concentrate exclusively on execution topology, while cognitive science studies focus on cognitive function, failing to distinguish architecturally unique systems. For instance, the Orchestrator-Workers topology can support various patterns like Plan-and-Execute or Hierarchical Delegation, each with unique failure modes. This paper can be found on arXiv under 2605.13850.