logseq/pages/@MODELS_2025_paper_19.md

tags:: #zotero title:: @MODELS_2025_paper_19 item-type:: document original-title:: MODELS_2025_paper_19 language:: en links:: Local library, Web library

• Attachments

  collapsed:: true
  • PDF {{zotero-imported-file 8XELL96A, "MODELS_2025_paper_19.pdf"}}

• Notes

  • I'm reviewing a research paper and I took the following notes:

• Annotazioni

  (5/5/2025, 16:40:59)

  • “Model-Driven Approach” (“MODELS_2025_paper_19”, p. 1) #ffd400 The paper makes an overview of different technologies for supporting multi agent systems and propose a taxonomy including and organizing different terms that are involved in the development of multi agents systems based on LLMs. As a such the title does not reflect the actual content of the paper, which is more a survey than an approach.

  • “LLM Agent Design” (“MODELS_2025_paper_19”, p. 1) #5fb236

  • “most agents are assembled in an ad-hoc manner without a coherent architectural foundation.” (“MODELS_2025_paper_19”, p. 1) #5fb236

  • “The challenge lies in structuring task decomposition, agent delegation, tool integration, and evaluation into coherent, extensible agent behaviors.” (“MODELS_2025_paper_19”, p. 1) #e56eee

  • “we propose a modeldriven framework for LLM agent design that distinguishes structural components from cognitive modules and integrates them into a unified architecture.” (“MODELS_2025_paper_19”, p. 1) #e56eee

  • “transforming how people interacted with LLMs and how developers began to architect intelligent capabilities atop existing and new software.” (“MODELS_2025_paper_19”, p. 1) #5fb236

  • “With the rise of frameworks like LangChain1 and LlamaIndex2, developers could integrate LLMs into applications independent of LLM vendor lock-in” (“MODELS_2025_paper_19”, p. 1) #5fb236

  • “from LLM workflows to agentic systems powered by LLMs capable of contextual, tool-mediated, and adaptive behavior.” (“MODELS_2025_paper_19”, p. 1) #e56eee

  • “these approaches were functionally and architecturally distinct from LLM-integrated software.” (“MODELS_2025_paper_19”, p. 1) #5fb236

  • “LLMs were used to classify model repositories (classification)” (“MODELS_2025_paper_19”, p. 1) #5fb236

  • “LLM agents—components that use LLMs not just to respond, but to plan, act, and adapt” (“MODELS_2025_paper_19”, p. 1) #e56eee

  • “OpenAI Swarm6, now replaced by Agents SDK7” (“MODELS_2025_paper_19”, p. 1) #5fb236

  • “reintroduces concepts from multi-agent systems and cognitive architectures,” (“MODELS_2025_paper_19”, p. 1) #e56eee

  • “the focus for developers moves beyond prompt writing and pipeline design to architectural and design choices involving planning modules, delegation logic, and tool interfaces.” (“MODELS_2025_paper_19”, p. 2) #e56eee THAT'S CRUCIAL! IMPORTANT. Maybe tool interfaces is related to MCP?

  • “architectural perspective” (“MODELS_2025_paper_19”, p. 2) #a28ae5

  • “agents” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “tasks” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “tools” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “orchestra” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “cognitive modules” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “architectural model for LLM agent design” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “systematized framework of structural components” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “integrated behavioral scaffold” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “Plan–Act–Evaluate” (“MODELS_2025_paper_19”, p. 2) #5fb236

  • “Explore–Exploit–Learn” (“MODELS_2025_paper_19”, p. 2) #5fb236

  • “Together, these contributions present a principled, modeldriven view of LLM-native systems, clarifying core elements and enabling deliberate design of orchestration, reasoning, and adaptability.” (“MODELS_2025_paper_19”, p. 2) #e56eee

  • “LLM usage to agentic integration” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “agent design architecture” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “structural components and cognitive modules” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “orchestration, decomposition, collaboration, and tool use” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “explore–exploit–learn dynamics” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “experimentation” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “future directions and conclusions” (“MODELS_2025_paper_19”, p. 2) #2ea8e5

  • “distributed autonomy and coordination” (“MODELS_2025_paper_19”, p. 2) #5fb236

  • “LLM-powered agents now exhibit memory, planning, and reasoning capabilities [18], with simulation pipelines embedding LLMs across their full lifecycle [19].” (“MODELS_2025_paper_19”, p. 2) #5fb236

  • “Frameworks have emerged that decompose agents into modules for planning, memory, tool use, and reflection [23],” (“MODELS_2025_paper_19”, p. 2) #e56eee

  • “showing convergence with MAS principles—but realized through LLM-native methods.” (“MODELS_2025_paper_19”, p. 2) #e56eee

  • “Although effective for interleaved reasoning and acting, these early agents lacked orchestration, memory-backed decision-making, or modular delegation logic—essential features for handling complex, long-horizon tasks. Without shared execution contexts or rolebased coordination, they could not generalize beyond isolated task flows.” (“MODELS_2025_paper_19”, p. 2) #5fb236

  • “foundational planning primitive embedded within structured architectures.” (“MODELS_2025_paper_19”, p. 3) #5fb236

  • “memory (short- and long-term),” (“MODELS_2025_paper_19”, p. 3) #a28ae5

  • “profile module” (“MODELS_2025_paper_19”, p. 3) #a28ae5

  • “adaptive planning” (“MODELS_2025_paper_19”, p. 3) #a28ae5

  • “Architectural Perspectives on LLM Agent Desig” (“MODELS_2025_paper_19”, p. 3) #2ea8e5

  • “Early frameworks introduced cognitive loops with modular memory, internal/external actions, and decision cycles [26]” (“MODELS_2025_paper_19”, p. 3) #e56eee

  • “while others organized agents around core functions of reasoning, perception, and action” (“MODELS_2025_paper_19”, p. 3) #e56eee

  • “Learning, Knowledge, and Adaptation in LLM Agents” (“MODELS_2025_paper_19”, p. 3) #2ea8e5

  • “As LLM agents gain traction, what remains missing is a cohesive, architectural model that ties these capabilities together with clarity and intentionality. Rather than introducing entirely new mechanisms, our work offers a model-driven perspective that unifies these components into a coherent design framework.” (“MODELS_2025_paper_19”, p. 3) #e56eee

  • “We believe this will enable developers to build context-aware, tool-mediated, and adaptive LLM agents with deliberate modularity and strategic coordination,” (“MODELS_2025_paper_19”, p. 3) #f19837

  • “The knowledge component supports semantic reasoning, and memory provides continuity and personalization across interactions.” (“MODELS_2025_paper_19”, p. 3) #a28ae5

  • “Task execution is governed by a modular orchestration layer that coordinates subtasks, tool usage, and inter-agent collaboration.” (“MODELS_2025_paper_19”, p. 3) #5fb236

  • “A. Operational Roles and Interfaces” (“MODELS_2025_paper_19”, p. 4) #5fb236

  • “It carries with it an evaluative dimension, as each goal is associated with specific criteria that define what success looks like.” (“MODELS_2025_paper_19”, p. 4) #5fb236

  • “This distinction is important because the same goal can be pursued through many different tasks, and tasks may need to be recomposed, reprioritized, or even reinterpreted as conditions change or new information arises. From a system perspective, this allows the agent to dynamically adjust its operational plan (tasks) while maintaining alignment with the overarching goal.” (“MODELS_2025_paper_19”, p. 4) #a28ae5

  • “The agent does not use tools indiscriminately; rather, tool invocation is governed by its internal capabilities, activated based on task requirements and environmental context” (“MODELS_2025_paper_19”, p. 4) #5fb236

  • “It maintains a shared context over the internal states of all agents, manages workflows, handles branching or fallback logic, and bridges perception, cognition, and action—enabling sustained, adaptive reasoning across stages and modalities.” (“MODELS_2025_paper_19”, p. 4) #5fb236

  • “Goal” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Sufficiency” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Preferences” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Guardrails” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Memory” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Learning” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Knowledge” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “epresentative types for each module” (“MODELS_2025_paper_19”, p. 5) #5fb236

  • “Goals” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Sufficiency” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Learning” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Guardrails” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “memory” (“MODELS_2025_paper_19”, p. 5) #2ea8e5

  • “Episodic memory” (“MODELS_2025_paper_19”, p. 5) #5fb236

  • “Reflective memory” (“MODELS_2025_paper_19”, p. 5) #5fb236

  • “Preference memory” (“MODELS_2025_paper_19”, p. 5) #5fb236

  • “main orchestration variations” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “task decomposition type” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “delegation and collaboration mechanisms” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “tool usage” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “orchestration architectures” (“MODELS_2025_paper_19”, p. 6) #5fb236

  • “publish-subscribe” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “graph-based” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “handoff-based” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “Publish-Subscribe” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “Graph-Based” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “Execution is centrally orchestrated based on graph logic, with a persistent shared Execution Context capturing state updates” (“MODELS_2025_paper_19”, p. 6) #5fb236

  • “Handoff-Based” (“MODELS_2025_paper_19”, p. 6) #2ea8e5

  • “Fig. 5.” (“MODELS_2025_paper_19”, p. 7) #ffd400 A legend with the used color schema is needed.

  • “As shown in Figure 7, a central agent interprets userprovided instructions and orchestrates the task through a mix of direct action, delegation to specialized multi-agent teams, and internal deliberation.” (“MODELS_2025_paper_19”, p. 7) #a28ae5

  • “A key mechanism in this setup is nested chat, which is an agentic version of prompt chaining. The nested chat can involve launching sub-conversations with other agents, running inner verification or refinement loops, or resolving ambiguities in the original instruction.” (“MODELS_2025_paper_19”, p. 7) #ffd400 How about governance or in general the resolution of possible conflicting situations?

  • “Figure 7 is just one instantiation; delegation and collaboration patterns can vary widely depending on the agentic framework and problem space.” (“MODELS_2025_paper_19”, p. 7) #5fb236

  • “Generic Tools” (“MODELS_2025_paper_19”, p. 8) #5fb236

  • “Function Tools” (“MODELS_2025_paper_19”, p. 8) #5fb236

  • “Agent-as-Tool” (“MODELS_2025_paper_19”, p. 8) #5fb236

  • “Agent Role Typology” (“MODELS_2025_paper_19”, p. 8) #2ea8e5

  • “exploreexploit-learn paradigm” (“MODELS_2025_paper_19”, p. 8) #5fb236

  • “Plan–Act–Evaluate (PAE)” (“MODELS_2025_paper_19”, p. 8) #2ea8e5

  • “Autogen, OpenAI Agents SDK, CrewAI, PydanticAI, and LangGraph” (“MODELS_2025_paper_19”, p. 9) #ffd400 What are the criteria that have been used to select them?

  • “12We note two concerns, beyond this paper’s scope: (a) open-source models (e.g., available via Ollama) are less capable than frontier models, often lacking tool use, vision, and reasoning, and tend to produce more erratic agent behavior; and (b) frontier model costs can become prohibitive as agents and messages proliferate.” (“MODELS_2025_paper_19”, p. 9) #e56eee

  • “hands-on experiences with agentic frameworks” (“MODELS_2025_paper_19”, p. 10) #5fb236

  • “design architecture across diverse domains to assess robustness, adaptability, and goal alignment in real-world settings.” (“MODELS_2025_paper_19”, p. 10) #5fb236

  • “benchmarks that evaluate orchestration quality” (“MODELS_2025_paper_19”, p. 10) #5fb236

  • “This paper responds to that shift by proposing a principled, model-driven framework for LLM agent design.” (“MODELS_2025_paper_19”, p. 10) #a28ae5

  • “we observed operational differences and early indications of recurring design tendencies” (“MODELS_2025_paper_19”, p. 10) #a28ae5

  • “set of practical guidelines for structuring agent behavior” (“MODELS_2025_paper_19”, p. 10) #a28ae5

  • “Rather than prescribing a monolithic architecture, we offer scaffolds that support flexible composition and architectural clarity.” (“MODELS_2025_paper_19”, p. 10) #ffd400 Indeed, I appreciate the effort, even though the proposed description is generic and vague. There are not specific guidelines on how to actually decompose monolithic architectures.

  • “REFERENCES” (“MODELS_2025_paper_19”, p. 11) #ffd400 34 references out of 50 come from arxive. I understand this is a very fast growing topic, however it is necessary to keep basing our research on peer reviewed sources. This is a general comment.

    COnsider that those that are tagget with #5fb236 are just highlights, those that are tagged with #e56eee and #a28ae5 are imporant sentences. Please pay attention instead to the notes that are tagged with #ffd400. Those that are tagged with #ff6666 are typos or errors. Could you please draft a review by organizing it as follows: SUMMARY: Just a few sentence to summarize the work COMMENTS: Organize the notes especially those that contain issues or typos.