Query Optimization Beyond Data Systems: The Case for Multi-Agent Systems

2025-12-10

AI Summary

Existing LLM-based multi-agent workflows lack a general, scalable query optimization mechanism for heterogeneous data sources and execution engines, resulting in high LLM invocation overhead, severe task redundancy, and inefficient collaborative reasoning. Method: This paper pioneers the adaptation of database query optimization paradigms to multi-agent systems, proposing the first query optimization framework specifically designed for multi-agent workflows. It integrates cost-aware modeling, workflow graph analysis, abstraction of heterogeneous executors, and LLM-invocation–aware optimization to enable automated model selection, cross-engine orchestration, and collaborative inference optimization. Contribution/Results: Experiments demonstrate that our framework significantly reduces LLM invocation costs (by 37% on average), eliminates redundant tasks, and improves execution efficiency for cross-engine workflows. This work establishes the foundational research paradigm of multi-agent query optimization and provides both theoretical grounding and architectural support for building low-cost, highly adaptive agent-based data pipelines.

Technology Category

• Search and Optimization: Learning to Search
• Multiagent Systems: Teamwork
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search

Abstract

The proliferation of large language models (LLMs) has accelerated the adoption of agent-based workflows, where multiple autonomous agents reason, invoke functions, and collaborate to compose complex data pipelines. However, current approaches to building such agentic architectures remain largely ad hoc, lacking generality, scalability, and systematic optimization. Existing systems often rely on fixed models and single execution engines and are unable to efficiently optimize multiple agents operating over heterogeneous data sources and query engines. This paper presents a vision for a next-generation query optimization framework tailored to multi-agent workflows. We argue that optimizing these workflows can benefit from redesigning query optimization principles to account for new challenges: orchestration of diverse agents, cost efficiency under expensive LLM calls and across heterogeneous engines, and redundancy across tasks. Led by a real-world example and building on an analysis of multi-agent workflows, we outline our envisioned architecture and the main research challenges of building a multi-agent query optimization framework, which aims at enabling automated model selection, workflow composition, and execution across heterogeneous engines. This vision establishes the groundwork for query optimization in emerging multi-agent architectures and opens up a set of future research directions.

Problem

• Optimizes multi-agent workflows for heterogeneous data sources
• Redesigns query optimization for LLM cost efficiency
• Enables automated model selection and workflow composition

Innovation

• Optimizing multi-agent workflows with query principles
• Automating model selection and workflow composition
• Enabling cross-engine execution for heterogeneous data