LN: Masterman et al. (2024) — The Landscape of Emerging AI Agent Architectures: A Survey

Bibliographic Reference

Citation: Masterman, T., Besen, S., Sawtell, M., & Chao, A. (2024). The landscape of emerging AI agent architectures for reasoning, planning, and tool calling: A survey. arXiv:2404.11584. https://arxiv.org/abs/2404.11584 (This paper is in the PUMA project knowledge PDFs as “The_Landscape_of_Emerging_AI_Agent_Architectures_for_Reasoning_Planning_and_Tool_Calling_A_Survey_2404.11584v1.pdf”)


Pass 1 — Bird’s Eye View (5 Cs)

CAssessment
CategorySurvey
ContextComprehensive survey of agent architectures post-ChatGPT era
CorrectnessCovers 30+ systems systematically. Well-organised taxonomy.
Contributions(1) Taxonomy: single-agent vs. multi-agent; (2) Reasoning patterns: CoT, ToT, ReAct, Reflexion; (3) Planning: task decomposition, tool selection; (4) Tool integration patterns
ClarityExcellent. Well-organised reference.

Relevance: ⭐⭐⭐⭐

Provides the theoretical taxonomy for classifying PUMA’s agent architecture choices. Essential reference for Ch.2 state-of-the-art section.


Pass 2 — Key Points

The survey’s taxonomy of single-agent vs. multi-agent architectures, and the spectrum from “tool-augmented LLM” to “fully autonomous agent,” provides the classification framework for positioning PUMA:

  • PUMA Stage 1–2: Tool-augmented LLM (Ollama inference + structured prompting)
  • PUMA Stage 4: Single agent with RAG tool (ReAct pattern)
  • PUMA Stage 5: Multi-agent with orchestration (MASAI/MetaGPT pattern)

PUMA Integration

  • Ch.2 (State of the art): Primary taxonomy reference for agent architecture classification
  • Architecture spec: SP-Architecture
  • PUMA stages: stages 1–5 map to autonomy levels in this survey → EX-Stages-Overview

MOCs