Optimal agentic LLM system design across use cases

Determine a clear, general approach for designing optimal agentic large language model (LLM) multi-agent systems across different application use cases, specifically establishing criteria and guidelines for selecting and orchestrating collaborating agents (e.g., Classifier, Retriever, Generator, Reviewer) within Generic Agentic RAG (GA-RAG) workflows to reduce hallucinations and improve task efficiency.

Background

The paper surveys multi-agent (agentic) LLM systems and notes that, despite promising results in various domains, there is no established methodology for choosing an optimal architecture for different tasks. To address this gap, the authors propose a Generic Agentic RAG (GA-RAG) workflow comprising Classifier, Retriever, Generator, and Reviewer agents, but they acknowledge that determining optimal design choices remains unresolved.

This open problem is motivated by the need to systematically specify how many agents to use, what roles they should play, and how they should be orchestrated to achieve reliable, efficient performance in knowledge-intensive, real-time retrieval-augmented generation systems for IoT and other domains.

References

This topic still needs further investigation, as we have no clear approach to determining the optimal design for different use cases.

— Agentic Search Engine for Real-Time IoT Data  (2503.12255 - Elewah et al., 15 Mar 2025) in Subsubsection "Agentic LLM System," Section "Background and Related Work"

Several directions remain open. The present ten-domain structure is a practical, expert-informed organization rather than an established optimum; its composition and coordination may need to adapt to different materials problems.

— Scaling LLM Agents for Materials Design through Hierarchical Collective Reasoning  (2609.16466 - Choi et al., 15 Sep 2026) in Conclusions, paragraph beginning “Several directions remain open”

These considerations lead to several open questions: What capabilities should be used to classify models so that appropriate models can be selected for different refactoring sub-tasks? What coordination mechanisms can resolve conflicts when multiple agents propose competing changes? How should the system architecture itself adapt as model capabilities evolve rapidly?

— Continuous Autonomous Refactoring: A Research Roadmap for AI-Driven Code Quality Maintenance  (2609.01236 - Sun et al., 1 Sep 2026) in Section 3.4, Architecture and Design Pattern Challenges