Efficient Tool Use in LLM Agents: A Survey of Bottlenecks and Optimization

PhD Qualifying Examination


Title: "Efficient Tool Use in LLM Agents: A Survey of Bottlenecks and 
Optimization"

by

Miss Jieling YU


Abstract:

The rapid evolution of large language models (LLMs) has transformed them from 
passive text generators into tool-augmented agents capable of interacting 
with external APIs, databases, execution environments, and distributed 
systems. As tool ecosystems continue to expand from function calling to Model 
Context Protocol (MCP) infrastructures and reusable skill libraries, modern 
agents increasingly face challenges in retrieval scalability, long-horizon 
planning, execution efficiency, and runtime coordination, making efficient 
tool use a critical bottleneck for next-generation agent systems.

This survey presents a systematic review of efficient tool use in LLM agents 
from both reasoning and systems perspectives. We first review the evolution 
of tool ecosystems, including APIs, MCP systems, and skill-based frameworks, 
and then organize existing agent workflows into a unified four-stage pipeline 
consisting of planning, retrieval, execution, and runtime orchestration. We 
further summarize representative benchmarks, major bottlenecks, and recent 
optimization techniques such as hierarchical retrieval, speculative planning, 
execution-aware scheduling, caching, and workflow memory. Our goal is to 
provide a structured understanding of scalable and efficient tool-augmented 
LLM agents and highlight promising directions for future research.


Date:                   Tuesday, 2 June 2026

Time:                   2:00pm - 3:00pm

Venue:                  Room 2129A
                        Lift 19

Committee Members:      Prof. Bo Li (Supervisor)
                        Dr. Wei Wang (Chairperson)
                        Dr. Xiaomin Ouyang