Coding agents are a type of AI-based automated software development system, typically centered around a Large Language Model (LLM). They assist or autonomously complete engineering tasks such as code generation, program debugging, defect fixing, and software maintenance by understanding natural language requirements, analyzing code repositories, invoking development tools, and executing testing tasks. Unlike traditional code completion tools, coding agents can plan multi-step processes around the target task and continuously adjust the execution process based on feedback from the external environment.

The development of coding agents is built upon large code models, agent frameworks, and software engineering automation techniques. In 2024, researchers from Princeton University and other institutions published a paper...*SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering*This paper proposes SWE-agent, an intelligent agent framework for software engineering tasks. The research introduces an Agent-Computer Interface (ACI), enabling large language models to interact with code repositories, terminal environments, and development tools through a specially designed interface. This allows them to perform code modifications, testing, and bug fixing in real-world GitHub software engineering tasks, providing an important reference for subsequent research on coding agents.

Encoding proxies are primarily used to address issues such as insufficient contextual understanding, lack of environmental feedback, and difficulty in handling complex engineering tasks when traditional large language models directly generate code. Their core processes typically include task understanding, code retrieval, modification plan planning, code editing, running tests, and iterative optimization based on feedback. By connecting compilers, terminals, version control systems, and testing frameworks, encoding proxies can form a closed-loop development process of "generation—execution—evaluation—correction."

In recent years, with the improvement of large language model code capabilities, coding agents have been increasingly applied to software development assistance scenarios. Representative tools include code editing and development assistants based on large models, such as Claude Code from Anthropic and Cursor Agent from Anysphere. These systems can help developers automate code writing, code understanding, debugging and analysis, and engineering tasks, but currently still require human supervision to ensure the correctness, security, and reliability of the generated code.