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

Jia Huang, Joey Tianyi ZhouZenodo 2026 · DOI

TL;DR

This paper introduces a novel two-dimensional framework for classifying AI agent architectures, combining "Cognitive Function" (what the agent does) and "Execution Topology" (how it's structured). This framework provides a much-needed, unified vocabulary to disambiguate and systematically design complex LLM-based agents, moving beyond existing single-axis approaches.

What Problem Does This Paper Address?

The rapid proliferation of LLM-based agent systems has led to a fragmented landscape of architectural guidance. Current frameworks tend to focus on either how data flows through an agent (execution topology, e.g., chains, orchestrators) or what the agent is capable of (cognitive function, e.g., reasoning, memory). However, this single-axis view is insufficient. For instance, the same "Orchestrator-Workers" topology can implement vastly different patterns like "Plan-and-Execute," "Hierarchical Delegation," or "Adversarial Verification," each with distinct failure modes and design trade-offs. This ambiguity makes it difficult to compare, design, and understand agent architectures effectively.

Methods

The authors propose a 7x6 matrix framework. The first axis, Cognitive Function, categorizes agent capabilities into seven types: Perception, Memory, Reasoning, Action, Reflection, Collaboration, and Governance. These are grounded in cognitive science and extended based on practical agent development. The second axis, Execution Topology, defines six structural archetypes: Chain, Route, Parallel, Orchestrate, Loop, and Hierarchy. These archetypes are designed to subsume existing topologies found in industry frameworks.

The combination of these two axes creates a 42-cell matrix, identifying 28 distinct agent design patterns. The paper details eight representative patterns, providing their coordinates in the matrix, problem statement, architectural solution, and engineering trade-offs. Orthogonality between the axes is demonstrated by showing that a single topology can serve multiple cognitive functions, and a single cognitive function can be realized by multiple topologies. The framework's descriptive power is validated through case studies across four real-world domains.

Datasets

Not discussed in the paper. The paper focuses on a conceptual framework and its application to existing domains rather than introducing new datasets for training or evaluation.

Key Insights and Ideas

  • The Orthogonality of Function and Topology is Crucial: The core insight is that understanding what an agent needs to do (cognitive function) and how it's structured (execution topology) are independent but equally vital dimensions for describing and designing AI agents.
  • Existing Frameworks are Incomplete: The paper convincingly argues that current industry and academic frameworks are limited by their single-axis perspective, leading to ambiguity and difficulty in comparing agent designs.
  • A Unified Vocabulary is Needed: The proposed 7x6 matrix provides a principled, framework-neutral, and model-agnostic vocabulary that can serve as a common language for AI agent architecture design.
  • Empirical Laws of Pattern Selection: By analyzing real-world domains, the authors derive five "laws" that link environmental constraints (like time pressure, action authority, failure cost asymmetry, and volume) to specific architectural choices, offering practical guidance for designers.
  • The Framework is Durable: The chosen categories for cognitive function and execution topology are designed to remain relevant as underlying LLM models and frameworks evolve.

Key Results

  • 28 Named Patterns: The framework identifies 28 distinct agent design patterns within the 7x6 matrix, with 15 original names coined in this work.
  • Coverage Across Domains: The framework successfully describes architectures across four diverse real-world domains (financial lending, legal due diligence, network operations, healthcare triage), demonstrating its broad applicability.
  • Five Empirical Laws: The study yields five invariant principles governing pattern selection, such as "Time pressure determines architectural complexity" and "Volume determines collaboration needs."
  • Framework-Neutral and Model-Agnostic: The proposed classification is designed to be independent of specific LLM models or agent frameworks, ensuring its long-term utility.

Limitations and Future Work

  • Non-Perfect Orthogonality: While aiming for orthogonality, some combinations of cognitive functions and execution topologies are more natural or common than others. The paper acknowledges that 28 out of 42 cells are populated, suggesting some combinations might be impossible, under-explored, or limited by current technology.
  • Pattern Granularity: The count of 28 patterns represents a specific level of granularity. The authors suggest that one could further subdivide patterns (e.g., Generator-Critic) or merge them, impacting the total count.
  • Tactical vs. Outer Execution Loop: The framework focuses on "tactical patterns" (bounded architectural primitives) and deliberately excludes the agent's outer execution loop (like the ReAct cycle), which is considered a runtime substrate rather than a peer pattern. This distinction is important for scope management.
  • Future Work: The authors invite the community to validate, extend, and challenge the framework. They also suggest that as agent systems mature, new patterns might emerge, particularly in under-explored cells of the matrix (e.g., Parallel Reflection, Reflection Orchestrate). An extended engineering treatment and implementation are also mentioned.

Summary updated

Abstract

Existing frameworks for LLM-based agent architectures describe systems from a single perspective. Industry guides such as those from Anthropic, Google, and LangChain emphasize execution topology—how information and control flow through the system—whereas cognitive-science-oriented surveys emphasize cognitive function—what the agent is trying to accomplish. We argue that neither axis alone is sufficient to distinguish architecturally different systems. For example, the same Orchestrator–Workers topology can realize Plan-and-Execute, Hierarchical Delegation, or Adversarial Verification, despite their substantially different failure modes and design trade-offs. We propose a two-dimensional classification framework that combines (1) a Cognitive Function axis with seven categories—Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, and Governance—and (2) an Execution Topology axis with six structural archetypes—Chain, Route, Parallel, Orchestrate, Loop, and Hierarchy. The resulting 7×6 matrix yields 27 distinct agent design patterns, including 13 patterns introduced and named in this work. We show that the two axes are orthogonal through systematic cross-axis analysis, present eight representative patterns in detail, and evaluate the descriptive coverage of the framework across four real-world domains: financial lending, legal due diligence, network operations, and healthcare triage. From this cross-domain analysis, we derive five empirical laws of pattern selection that relate environmental constraints—such as time pressure, action authority, failure-cost asymmetry, and task volume—to architectural choices. The framework provides a principled, framework-neutral, and model-agnostic vocabulary for designing and comparing AI agent architectures.
BibTeX
@inproceedings{huang2026two,
  author = {Jia Huang and Joey Tianyi Zhou},
  title = {{A Two-Dimensional Framework for AI Agent Design Patterns: Cognitive Function × Execution Topology}},
  booktitle = {Zenodo},
  year = {2026},
  doi = {10.5281/zenodo.19036556},
}

Loading related content...

Figure