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

Jia Huang; Joey Tianyi Zhou

cs.AIMar 16, 2026

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

Authors: Jia Huang, Joey Tianyi Zhou

Organizations: Agency for Science, Technology and Research (A*STAR), Singapore · Centre for Frontier AI Research (CFAR), A*STAR

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 (Perception, 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 28 named patterns, 15 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.

Figures & tables

ID	Category	Core Question
C1	Perception	What information enters the agent’s working memory?
C2	Memory	How does the agent store, retrieve, and update knowledge?
C3	Reasoning	How does the agent deliberate and decide?
C4	Action	How does the agent act on the world through tools?
C5	Reflection	How does the agent evaluate and improve its own outputs?
C6	Collaboration	How do multiple agents coordinate to solve a problem?

Table 1: Seven cognitive function categories.

ID	Archetype	Structure
T1	Chain	Linear sequential pipeline; output of step $n$ feeds step $n+1$
T2	Route	Conditional branching; a classifier dispatches to specialized handlers
T3	Parallel	Concurrent fan-out with aggregation; independent subtasks run simultaneously
T4	Orchestrate	Central coordinator delegates to workers and synthesizes results
T5	Loop	Iterative refinement with explicit exit conditions
T6	Hierarchy	Nested multi-level delegation; each level can use any other topology

Table 2: Six execution topology archetypes.

	T1 Chain	T2 Route	T3 Parallel	T4 Orchestrate	T5 Loop	T6 Hierarchy
C1	Semantic Compact. $\bigstar$	Context Triage $\bigstar$	Multi-Modal Fusion	Progressive Disc. $\bigstar$	—	—
C2	RAG Pipeline	Hierarchical Ret. $\bigstar$	—	Progress Track. $\bigstar$	Failure Journal $\bigstar$	—
C3	Chain-of-Thought	Complexity Rte. $\bigstar$	Parallel Explor.	—	Iterative Hyp. $\bigstar$	—
C4	Prompt Chaining	Tool Dispatch	—	Plan-and-Execute	—	Guardrail Sand. $\bigstar$
C5	Generator-Critic	Skill Package $\bigstar$	—	—	Self-Heal Loop $\bigstar$	Exp. Replay
C6	Handoff Chain	—	Fan-Out/Gather	—	Adversarial Rev.	Hier. Deleg.

Table 3: The

7\times 6

pattern matrix. 28 named patterns;

\bigstar

= original name coined in this work.

Resource	Cognitive Axis?	Topology Axis?	# Patterns	Framework-Neutral?
Anthropic [ 1 ]	No	Yes (6)	6	Yes
Google ADK [ 2 ]	No	Yes (8)	8	No
LangChain [ 3 ]	No	Yes (4)	4	No
Andrew Ng [ 4 ]	Yes (4)	No	4	Yes
Wang et al. [ 5 ]	Yes (4)	No	—	Yes
Sumers et al. [ 7 ]	Yes (4)	No	—	Yes

Table 4: Comparison with existing agent architecture resources.

	Financial Lending	Legal Due Diligence	Network Operations	Healthcare Triage
Task	SME loan assessment	M&A contract review	NOC alert handling	ED patient triage
Time budget	4 hours	8 hours	5 minutes	60 seconds
Volume	1 case	500 contracts	Continuous stream	1 patient
Primary topology	Orchestrate	Hierarchy	Route	Chain
Patterns selected	7	8	9	7
Action authority	Recommend only	Recommend only	Auto (P3/P4)	Recommend only

Table 5: Four case study domains with structurally different architectures derived from the same pattern catalog.

AI Agent Architecture

See all 25 papers

A Methodology for Selecting and Composing Runtime Architecture Patterns for Production LLM Agents
May 19, 2026Vasundra SrinivasanLLM Agent Orchestration LLM Agents
AI Architect, Independent Researcher · Stanford School of Engineering
Architectural Design Decisions in AI Agent Harnesses
Apr 20, 2026Hu WeiLLM Agent Orchestration Tool-Using Agents
Architectural Design, Not Only Model Intelligence, Governs Multi-Agent LLM Performance
Feb 3, 2026Abdelghny Orogat, Ana Rostam, Essam MansourMulti-Agent LLM Systems Multi-Agent Orchestration
Concordia University

May 19, 2026J/K move · Enter open · S save

A Methodology for Selecting and Composing Runtime Architecture Patterns for Production LLM Agents

Vasundra Srinivasan

AI Architect, Independent Researcher · Stanford School of Engineering

LLM Agent Orchestration LLM Agents Long-Horizon LLM Agents LLM Agent Reliability

Open paper AskPDF arXiv

Production LLM agents combine stochastic model outputs with deterministic software systems, yet the boundary between the two is rarely treated as a first-class architectural object. This paper names that boundary the stochastic-deterministic boundary (SDB): a four-part contract among a proposer, verifier, commit step, and reject signal that specifies how an LLM output becomes a system action. We argue that the SDB is the load-bearing primitive of production agent runtimes. Around this primitive, we organize agent runtime design into three concerns: Coordination, State, and Control. We present a catalog of six runtime patterns that compose the SDB differently across conversational, autonomous, and long-horizon agents: hierarchical delegation, scatter-gather plus saga, event-driven sequencing, shared state machine, supervisor plus gate, and human in the loop. For each pattern, we trace its lineage to distributed-systems concepts and identify what changes when the worker is stochastic. The paper contributes a five-step methodology for selecting runtime patterns, a diagnostic procedure that maps production failures to pattern weaknesses, and a failure mode called replay divergence, in which LLM-based consumers of a deterministic event log produce different downstream outputs under model-version or prompt changes. A stylized reliability decomposition separates per-call model variance from architectural momentum, motivating the claim that as model variance decreases, pattern choice and SDB strength become increasingly important levers for long-run reliability. We apply the methodology to five workloads and provide one runnable reference implementation for a 90-day contract-renewal agent.

Abstract

AI Agent Architecture

Explore similar work

A Methodology for Selecting and Composing Runtime Architecture Patterns for Production LLM Agents

Architectural Design Decisions in AI Agent Harnesses

Architectural Design, Not Only Model Intelligence, Governs Multi-Agent LLM Performance