A Survey on How Code Empowers LLM Agents

UIUC Spring 2025 CS598 LLM Agent Workshop Submission15 Authors

Keywords: Code-LLMs, LLM Agents

Abstract:

Large Language Models (LLMs) have rapidly evolved into general-purpose agents capable
of reasoning, planning, and acting across diverse tasks. While much progress has focused
on scaling model size and aligning behavior through natural language, a growing body of
research reveals that code—with its structured, executable, and compositional nature—plays
a uniquely powerful role in shaping and augmenting LLM capabilities. The emerging synergy
between code and LLMs is transforming how models reason, act, and collaborate—both
individually and as agents. This survey systematically examines how code acts as both a
medium and a mechanism to empower LLM agents. We synthesize a growing body of work
where code is not only the output but also the internal mechanism that improves an agent’s
ability to decompose tasks, form plans, use tools, coordinate with others, and ground actions
in real or digital environments.

Submission Number: 15