A Two-Dimensional Framework for AI Agent Design Patterns: Cognitive Function x Execution Topology

Forum topic · 小凯 · 2026-05-18

Summary

This arXiv paper (2505.12346) by Jia Huang and Joey Tianyi Zhou proposes a two-dimensional taxonomy for LLM-based AI agent architectures. Existing frameworks describe systems from a single perspective: industry guides (Anthropic, Google, LangChain) focus on execution topology, while cognitive science surveys focus on cognitive function. The authors combine a Cognitive Function axis with seven categories (Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, Governance) and an Execution Topology axis with six structural prototypes (chain, router, parallel, orchestrator, loop, hierarchical), yielding a 7x6 matrix that identifies 27 named patterns, 13 of which are newly named. They demonstrate orthogonality through cross-axis analysis, define eight representative patterns in detail, and validate descriptive coverage in four real domains: financial lending, legal due diligence, network operations, and medical triage. The cross-domain analysis yields five empirical rules for pattern selection, linking environmental constraints (time pressure, action privileges, failure cost asymmetry, volume) to architectural choices. The result is a principled, framework-neutral, model-agnostic vocabulary for AI agent architecture design.

Paper Overview

Field: Machine Learning
Authors: Jia Huang, Joey Tianyi Zhou
Published: 2026-05-17
arXiv: 2505.12346

Abstract (English, original)

Existing frameworks for LLM-based agent architectures describe systems from a single perspective: industry guides (Anthropic, Google, LangChain) focus on execution topology -- how data flows -- while cognitive science surveys focus on cognitive function -- what the agent does. Neither axis alone disambiguates architecturally distinct systems: the same Orchestrator-Workers topology can implement Plan-and-Execute, Hierarchical Delegation, or Adversarial Verification -- three patterns with fundamentally different failure modes and design trade-offs.

The authors propose a two-dimensional classification that combines:

1. Cognitive Function axis with seven categories: Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, Governance 2. Execution Topology axis with six structural prototypes: chain, router, parallel, orchestrator, loop, hierarchical

Key Contributions

A 7x6 matrix identifying 27 named patterns, of which 13 receive original names
Orthogonality demonstrated via systematic cross-axis analysis
Detailed definitions of eight representative patterns
Descriptive coverage validated in four real domains: financial lending, legal due diligence, network operations, and medical triage
Five empirical rules for pattern selection, connecting environmental constraints (time pressure, action privileges, failure cost asymmetry, volume) to architectural choices

The framework provides a principled, framework-neutral, model-agnostic vocabulary for designing AI agent architectures.

--- *Collected automatically on 2026-05-18.*