A review of code agents based on large language models
Year: 2026
Type: article
Abstract: With the rapid advancement of large language models (LLMs), code agents have emerged as a key paradigm for intelligent and automated software development. While single-agent systems demonstrate strong capabilities in code generation, debugging, and testing, they struggle with complex, multi-stage software engineering workflows due to limitations in context length, task decomposition, and long-term memory. To overcome these bottlenecks, researchers have introduced multi-code agent systems, in which multiple agents with specialized roles collaborate through shared memory, role division, and collective reflection. This architecture enables more robust reasoning, coordinated planning, and scalable code generation in real-world projects. This paper provides a comprehensive review of the evolution from single-agent to multi-agent frameworks, analyzing core advancements in task planning, tool integration, memory management, and collaborative optimization. It further summarizes representative studies and application scenarios in software development automation. Finally, the paper discusses ongoing challenges-such as maintaining contextual consistency, optimizing multi-agent coordination, and ensuring code safety-and explores future directions toward self-evolving and industry-level intelligent programming systems.
Language: Chinese
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162354219