Instructions as code – instructing agentic AI models for changes in large projects

Software developers and companies are increasingly adopting agentic AI to implement software features. In practice, developers guide AI agents through explicit instructions that specify, for example, how to navigate the project, identify relevant source files, and test the generated code. Without such instructions, agents may blindly perform implementations leading to incomplete or incorrect results.

While instructions are intended to support AI-assisted development, recent work shows that instructions alone are insufficient to effectively guide agentic systems. Instead, instruction quality and level of detail play a critical role in agent performance. Despite a large body of research on the use of AI and large language models (LLMs) in software engineering, most existing studies focus on prompt-engineering techniques. However, recent findings suggest that prompt engineering provides limited benefits for advanced models, particularly those with strong reasoning capabilities. As a result, developers increasingly lack guidance on how to write high-quality instructions, especially in complex software projects involving multiple interacting components and third-party dependencies.

In this project, we aim to study how developers formulate instructions in relation to software system complexity, characterized by interconnected components and external dependencies, with the goal of deriving guidance for writing effective instructions for agentic AI in real-world software projects.

Faculty Supervisor:

Mohammed Sayagh

Student:

Partner:

Nara Institute of Science and Technology

Discipline:

Engineering

Sector:

Technology; Artificial Intelligence

University:

École de technologie supérieure

Program: