Abstract
The development of large language models has enabled AI-assisted programming to gradually expand from early code completion to requirements understanding, code generation, testing, debugging, and project maintenance. The programming intelligence agent formed on this basis is no longer just returning a piece of code based on user input, but can read project files, break down tasks, call terminals and testing tools, and repeatedly modify the program based on the running results. Based on 60 publicly searchable representative papers, this paper adopts a structured narrative review method to review the research progress from the aspects of code-based models, agent control mechanisms, repository-level systems, evaluation benchmarks, and human-computer collaboration evidence. It also compares the research scope and analysis focus with existing reviews on software engineering agents. This paper further utilizes method classification, performance statistics, and mechanism diagrams to analyze different technical approaches, focusing on issues such as reliability, contextual understanding in large-scale projects, security and privacy, evaluation and reproducibility, operating costs, and accountability and governance. Research shows that programming agents can reduce the cost of repetitive coding and help users lacking professional programming experience complete prototype development. However, its efficiency gains are significantly affected by task size, project familiarity, quality standards, and manual review costs, making it difficult to consistently understand requirements and guarantee overall quality in complex projects. Future research should shift from simply pursuing the quantity of code generated to building verifiable, explainable, traceable, and collaborative software development processes.