Programming with an AI Agent

Artificial intelligence (AI) has fundamentally changed the way we program. AI agents can generate, optimize, and even assist with debugging code. However, there are some limitations that programmers should keep in mind when working with AI.

Problems with order and duplication

AI agents have difficulty with the correct order of code. For example, they might place initializations at the end of a file, causing runtime errors. Additionally, AI can readily define multiple versions of the same class or function within a project, leading to conflicts and confusion.

A code platform with memory and project structure helps

One solution is to use AI code platforms that can manage memory and project structures. This helps maintain consistency in complex projects. Unfortunately, these features are not always applied consistently. As a result, the AI may lose the coherence of a project and introduce unwanted duplications or incorrect dependencies during programming.

Most AI coding platforms work with so-called tools that the large language model can call upon. These tools are based on an open standard protocol (MCP). It is therefore possible to link an AI coding agent to an IDE such as Visual Studio Code. Optionally, you can set up an LLM locally with llama or ollama and choose an MCP server to integrate with. Models can be found on huggingface.

IDE extensions are indispensable

To better manage AI-generated code, developers can use IDE extensions that ensure code correctness. Tools like linters, type checkers, and advanced code analysis tools help detect and correct errors early on. They are an essential complement to AI-generated code to ensure quality and stability.

The cause of recurring errors: context and role in APIs

One of the main reasons AI agents continue to repeat errors lies in how AI interprets APIs. AI models need context and a clear role description to generate effective code. This means prompts must be complete: they should not only include functional requirements but also explicitly state the expected outcome and the boundary conditions. To facilitate this, you can save prompts in a standard format (MDC) and send them to the AI by default. This is particularly useful for generic programming rules you adhere to, as well as the functional and technical requirements and the structure of your project.

Tools like FAISS and LangChain help

Products like FAISS and LangChain offer solutions to help AI better handle context. FAISS, for example, aids in efficiently searching and retrieving relevant code snippets, while LangChain helps structure AI-generated code and maintain context within a larger project. However, here too, you can optionally set it up locally with vector databases.

Conclusion: useful, but not yet autonomous

AI is a powerful tool for programmers and can help accelerate development processes. However, it is not yet truly capable of independently designing and building a more complex codebase without human oversight. Programmers should view AI as an assistant that can automate tasks and generate ideas, but which still requires guidance and correction to achieve a good result.

Contact us to help set up the development environment to assist teams in getting the most out of their development environment, allowing them to focus more on requirements engineering and design than on debugging and writing code.