Paper Overview
- Field: Machine Learning
- Authors: Jia Huang, Joey Tianyi Zhou
- arXiv: 2505.12346
Abstract (translated)
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 taxonomy combining:
1. Cognitive Function axis with seven categories: Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, and Governance 2. Execution Topology axis with six structural prototypes: Chain, Routing, Parallel, Orchestrator, Loop, and Hierarchical
The resulting 7×6 matrix identifies 27 named patterns, of which 13 carry original names. The paper demonstrates orthogonality through systematic cross-axis analysis, provides detailed definitions of eight representative patterns, and validates descriptive coverage in four real-world domains: financial lending, legal due diligence, network operations, and medical triage.
Cross-domain analysis yields five empirical rules for pattern selection, governing the relationship between environmental constraints (time pressure, action permissions, failure cost asymmetry, volume) and architectural choices.
The framework offers a principled, framework-neutral, and model-independent vocabulary for AI agent architecture design.
---
*Auto-collected on 2026-05-18*