arxiv: 2605.13850 · v1 · submitted 2026-03-16 · 💻 cs.AI · cs.MA· cs.SE

Recognition: no theorem link

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

Jia Huang , Joey Tianyi Zhou

Authors on Pith no claims yet

Pith reviewed 2026-05-15 11:00 UTC · model grok-4.3

classification 💻 cs.AI cs.MAcs.SE

keywords AI agent design patternsLLM agent architecturescognitive function classificationexecution topologypattern selection lawsagent frameworktwo-dimensional classification

0 comments

The pith

A 7-by-6 matrix of cognitive functions and execution topologies classifies 27 distinct AI agent design patterns.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that single-axis descriptions of LLM-based agents leave architecturally different systems indistinguishable. Crossing seven cognitive-function categories with six execution-topology archetypes creates a grid in which patterns that share the same data flow can still be separated by their distinct failure modes and trade-offs. Validation across financial lending, legal due diligence, network operations and healthcare triage shows the grid covers real systems, while cross-domain comparison produces five empirical laws that tie environmental constraints to preferred architectural choices.

Core claim

Combining a Cognitive Function axis of Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration and Governance with an Execution Topology axis of Chain, Route, Parallel, Orchestrate, Loop and Hierarchy produces a 7x6 matrix that identifies 27 named patterns, 13 of them original. Systematic cross-axis analysis establishes orthogonality between the two dimensions. Eight representative patterns are defined in detail, and analysis of four application domains yields five empirical laws relating constraints such as time pressure, action authority, failure cost asymmetry and volume to pattern selection.

What carries the argument

The 7x6 matrix formed by crossing seven cognitive-function categories with six execution-topology archetypes, which separates architecturally distinct agent systems even when they share the same execution structure.

If this is right

Agents sharing the same execution topology can implement different cognitive functions and therefore exhibit different failure modes.
Five empirical laws link measurable environmental constraints to the selection of specific patterns.
The resulting vocabulary is independent of any particular model or implementation framework.
The matrix provides descriptive coverage for agents operating in financial, legal, network and healthcare settings.
Eight patterns are elaborated with sufficient detail to support direct engineering use.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Design teams could treat the matrix as a checklist when choosing or auditing an agent architecture.
Mismatch between an observed pattern and the constraints of its deployment environment could be flagged as a source of elevated risk.
Quantitative performance data collected per cell could test and refine the five empirical laws.
Mapping additional domains would reveal whether the laws generalize or require new categories.

Load-bearing premise

The seven cognitive categories and six topology archetypes are exhaustive and mutually orthogonal enough to place every distinct agent architecture into exactly one cell of the matrix.

What would settle it

Discovery of an agent architecture that cannot be assigned to a single unique cell without overlap or omission, or of two differently named patterns that display identical failure modes and design trade-offs under the same constraints.

read the original abstract

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. We propose a two-dimensional classification that combines (1) a Cognitive Function axis with seven categories (Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, Governance) and (2) an Execution Topology axis with six structural archetypes (Chain, Route, Parallel, Orchestrate, Loop, Hierarchy). The resulting 7x6 matrix identifies 27 named patterns, 13 with original names. We demonstrate orthogonality through systematic cross-axis analysis, define eight representative patterns in detail, and validate descriptive coverage across four real-world domains (financial lending, legal due diligence, network operations, healthcare triage). Cross-domain analysis yields five empirical laws of pattern selection governing the relationship between environmental constraints (time pressure, action authority, failure cost asymmetry, volume) and architectural choices. The framework provides a principled, framework-neutral, and model-agnostic vocabulary for AI agent architecture design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a two-dimensional framework for LLM-based AI agent design patterns that combines a Cognitive Function axis (seven categories: Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, Governance) with an Execution Topology axis (six archetypes: Chain, Route, Parallel, Orchestrate, Loop, Hierarchy). The resulting 7x6 matrix yields 27 named patterns (13 original), demonstrates orthogonality via cross-axis analysis, details eight representative patterns, validates descriptive coverage across four domains (financial lending, legal due diligence, network operations, healthcare triage), and derives five empirical laws of pattern selection from environmental constraints such as time pressure and failure cost asymmetry.

Significance. If the taxonomy and orthogonality hold, the framework supplies a model-agnostic, framework-neutral vocabulary that disambiguates architecturally distinct systems better than single-axis industry or cognitive-science guides, with concrete examples and domain coverage that could guide design trade-offs; the five laws, if generalizable, would link constraints to architectural choices in a falsifiable way.

major comments (2)

[Cross-domain analysis] Cross-domain analysis section: the five empirical laws are derived solely from qualitative interpretive mapping of observed architectures onto the 7x6 matrix across four domains; without quantitative metrics, a falsification protocol, or tests on additional domains, they risk capturing domain-specific regularities rather than general principles.
[§3 (Framework Definition)] §3 (Framework Definition): the choice of exactly seven cognitive-function categories and six execution-topology archetypes is presented without an independent argument or external validation for exhaustiveness and mutual orthogonality beyond the taxonomy's own definitions, which is load-bearing for the claim of 27 distinct patterns.

minor comments (2)

[Results] A summary table listing all 27 patterns with their axis coordinates would improve readability and allow readers to verify the matrix coverage directly.
[Orthogonality analysis] The orthogonality demonstration relies on systematic cross-tabulation; adding a brief discussion of potential edge cases or overlapping patterns would strengthen the claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive feedback on our manuscript arXiv:2605.13850. We have carefully considered each major comment and provide point-by-point responses below. The revisions we propose will strengthen the presentation of the framework's foundations and the status of the empirical laws while preserving the core contributions.

read point-by-point responses

Referee: [Cross-domain analysis] Cross-domain analysis section: the five empirical laws are derived solely from qualitative interpretive mapping of observed architectures onto the 7x6 matrix across four domains; without quantitative metrics, a falsification protocol, or tests on additional domains, they risk capturing domain-specific regularities rather than general principles.

Authors: We agree that the five empirical laws are derived from qualitative interpretive mapping across the four domains and therefore constitute initial empirical observations rather than fully validated general principles. In the revision we will expand the Cross-domain analysis section to explicitly label the laws as hypotheses, add a dedicated Limitations and Future Validation subsection that outlines a falsification protocol (including quantitative metrics such as task-success rates under controlled constraint variations and tests on at least two additional domains), and temper the language to reflect the current scope. This directly addresses the concern about domain-specific regularities while retaining the descriptive utility of the existing analysis. revision: partial
Referee: [§3 (Framework Definition)] §3 (Framework Definition): the choice of exactly seven cognitive-function categories and six execution-topology archetypes is presented without an independent argument or external validation for exhaustiveness and mutual orthogonality beyond the taxonomy's own definitions, which is load-bearing for the claim of 27 distinct patterns.

Authors: The seven cognitive functions and six topologies were synthesized from a review of cognitive architectures and multi-agent systems literature (detailed in the Related Work section). To meet the referee's concern we will revise §3 by inserting a new subsection that states the explicit selection criteria (coverage of core agent capabilities and structural patterns identified in prior taxonomies) and demonstrates orthogonality via the cross-axis examples already present, where identical topologies paired with different functions produce distinct failure modes and design trade-offs. We will also note that full exhaustiveness remains an empirical claim subject to future refinement rather than an a priori proof. revision: yes

Circularity Check

0 steps flagged

No significant circularity: taxonomy and laws are constructed directly from defined axes without reduction to fits or self-citations.

full rationale

The paper defines a Cognitive Function axis (seven categories) and Execution Topology axis (six archetypes) as independent inputs, then populates the resulting 7x6 matrix to name 27 patterns. Orthogonality is shown via explicit cross-axis enumeration, representative patterns are defined by combining the axes, and the five empirical laws are stated as observations from mapping four external domains onto the matrix. No equation or claim reduces by construction to a fitted parameter, renamed input, or load-bearing self-citation; the framework remains a descriptive classification whose central outputs are generated from the stated definitions rather than presupposed by them.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework introduces no free parameters or new physical entities. It rests on two domain assumptions about the chosen categories being adequate to partition agent designs.

axioms (2)

domain assumption The seven cognitive functions (Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, Governance) form an exhaustive and mutually exclusive set for describing agent cognition.
Invoked when defining the Cognitive Function axis and the resulting matrix.
domain assumption The six structural archetypes (Chain, Route, Parallel, Orchestrate, Loop, Hierarchy) capture all relevant execution topologies.
Invoked when defining the Execution Topology axis and the resulting matrix.

pith-pipeline@v0.9.0 · 5543 in / 1629 out tokens · 53268 ms · 2026-05-15T11:00:11.023341+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 4 internal anchors

[1]

Building effective agents,

E. Schluntz and B. Zhang, “Building effective agents,” Anthropic Research Blog, Dec. 2024

work page 2024
[2]

Agent Development Kit: A flexible framework for building multi-agent systems,

Google Cloud, “Agent Development Kit: A flexible framework for building multi-agent systems,” Google Developers Blog, Apr. 2025

work page 2025
[3]

LangGraph: Multi-agent workflows,

H. Chase et al., “LangGraph: Multi-agent workflows,” LangChain Documentation, Feb. 2025

work page 2025
[4]

What’s next for AI agentic workflows,

A. Ng, “What’s next for AI agentic workflows,” Sequoia Capital AI Ascent, Mar. 2024

work page 2024
[5]

A survey on large language model based autonomous agents,

L. Wang, C. Ma, X. Feng, et al., “A survey on large language model based autonomous agents,”Frontiers of Computer Science, vol. 18, art. 186345, 2024

work page 2024
[6]

The Rise and Potential of Large Language Model Based Agents: A Survey

Z. Xi, W. Chen, X. Guo, et al., “The rise and potential of large language model based agents: A survey,” arXiv:2309.07864, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[7]

Cognitive architectures for language agents,

T. R. Sumers, S. Yao, K. Narasimhan, and T. L. Griffiths, “Cognitive architectures for language agents,”Foundations and Trends in Machine Learning, vol. 17, no. 6, pp. 882– 971, 2024. 10

work page 2024
[8]

Gamma, R

E. Gamma, R. Helm, R. Johnson, and J. Vlissides,Design Patterns: Elements of Reusable Object-Oriented Software. Addison-Wesley, 1994

work page 1994
[9]

Fowler,Patterns of Enterprise Application Architecture

M. Fowler,Patterns of Enterprise Application Architecture. Addison-Wesley, 2002

work page 2002
[10]

Agent design pattern catalogue: A collection of architectural patterns for foundation model based agents,

Y. Liu, S. K. Lo, Q. Lu, et al., “Agent design pattern catalogue: A collection of architectural patterns for foundation model based agents,”J. Systems and Software, vol. 220, art. 112278, 2024

work page 2024
[11]

Agentic design patterns: A system-theoretic framework,

M.-D. Dao, Q. M. Le, H. T. Lam, et al., “Agentic design patterns: A system-theoretic framework,” arXiv:2601.19752, 2026

work page arXiv 2026
[12]

Lost in the middle: How language models use long contexts,

N. F. Liu, K. Lin, J. Hewitt, et al., “Lost in the middle: How language models use long contexts,”Transactions of the Association for Computational Linguistics, vol. 12, pp. 157– 173, 2024

work page 2024
[13]

Retrieval-augmented generation for knowledge- intensive NLP tasks,

P. Lewis, E. Perez, A. Piktus, et al., “Retrieval-augmented generation for knowledge- intensive NLP tasks,”Advances in Neural Information Processing Systems, vol. 33, pp. 9459–9474, 2020

work page 2020
[14]

MemGPT: Towards LLMs as Operating Systems

C. Packer, S. Wooders, K. Lin, et al., “MemGPT: Towards LLMs as operating systems,” arXiv:2310.08560, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[15]

Kahneman,Thinking, Fast and Slow

D. Kahneman,Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011

work page 2011
[16]

RouteLLM: Learning to Route LLMs with Preference Data

I. Ong, A. Almahairi, V. Wu, et al., “RouteLLM: Learning to route LLMs with preference data,” arXiv:2406.18665, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[17]

Garcia-Molina and K

H. Garcia-Molina and K. Salem, “Sagas,”ACM SIGMOD Record, vol. 16, no. 3, pp. 249– 259, 1987

work page 1987
[18]

Large language models cannot self-correct reasoning yet,

J. Huang, X. Chen, S. Mishra, et al., “Large language models cannot self-correct reasoning yet,”Proc. ICLR, 2024

work page 2024
[19]

CRITIC: Large language models can self-correct with tool-interactive critiquing,

Z. Gou, Z. Shao, Y. Gong, et al., “CRITIC: Large language models can self-correct with tool-interactive critiquing,”Proc. ICLR, 2024

work page 2024
[20]

Self-Refine: Iterative refinement with self- feedback,

A. Madaan, N. Tandon, P. Gupta, et al., “Self-Refine: Iterative refinement with self- feedback,”Advances in Neural Information Processing Systems, vol. 36, 2023

work page 2023
[21]

Improving Factuality and Reasoning in Language Models through Multiagent Debate

Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch, “Improving factual- ity and reasoning in language models through multiagent debate,”Proc. ICML, 2024. arXiv:2305.14325

work page internal anchor Pith review Pith/arXiv arXiv 2024
[22]

ReAct: Synergizing reasoning and acting in language models,

S. Yao, J. Zhao, D. Yu, et al., “ReAct: Synergizing reasoning and acting in language models,”Proc. ICLR, 2023

work page 2023
[23]

Chain-of-thought prompting elicits reasoning in large language models,

J. Wei, X. Wang, D. Schuurmans, et al., “Chain-of-thought prompting elicits reasoning in large language models,”Advances in Neural Information Processing Systems, vol. 35, pp. 24824–24837, 2022

work page 2022
[24]

Hohpe and B

G. Hohpe and B. Woolf,Enterprise Integration Patterns: Designing, Building, and Deploy- ing Messaging Solutions. Addison-Wesley, 2003

work page 2003
[25]

Huang,Designing AI Agents: Patterns for Building Intelligent Systems

J. Huang,Designing AI Agents: Patterns for Building Intelligent Systems. Manning Pub- lications, 2026 (in preparation). 11

work page 2026