I. Overview of Coding Agents

A coding agent is an AI-based automated software development system that typically centers on Large Language Models (LLMs). By understanding natural language requirements, analyzing codebases, invoking development tools, and executing test tasks, these agents assist in or autonomously complete engineering tasks such as code generation, program debugging, defect repair, and software maintenance. Unlike traditional code completion tools, coding agents can perform multi-step planning around specific goals and continuously adjust their execution processes based on external environmental feedback.

The development of coding agents rests upon the foundations of large models for code, agent frameworks, and software engineering automation technologies. In 2024, researchers from Princeton University and other institutions published the paper *SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering, proposing the SWE-agent framework designed for software engineering tasks. This research introduced the "Agent-Computer Interface" (ACI), enabling large language models to interact with code repositories, terminal environments, and development tools through specially designed operational interfaces. It executed code modifications, testing, and issue resolution in real-world GitHub software engineering projects, providing significant reference points for subsequent coding agent research.

II. Core Functionality and Workflow

Coding agents primarily address issues inherent when using traditional large language models to generate code directly, such as insufficient context comprehension, lack of environment feedback, and difficulty handling complex engineering tasks. Their core workflow generally includes task understanding, code retrieval, modification plan formulation, code editing, runtime testing, and iterative optimization based on feedback. By connecting compilers, terminals, version control systems, and testing frameworks, coding agents form a closed-loop development process characterized by "generate–execute–evaluate–correct."

In recent years, driven by improvements in the coding capabilities of large language models, coding agents have gradually been applied to auxiliary scenarios in software development. Representative tools include model-based code editors and development assistants such as Anthropic's Claude Code and Anysphere's Cursor Agent. Such systems help developers accomplish code writing, code comprehension, debug analysis, and engineering task automation; however, human supervision remains necessary at present to ensure the correctness, security, and engineering reliability of generated code.