The author conducted experiments using LLM agents to develop customizations for a closed PLM system. The process involved providing agents with context, API documentation, and data schemas to automate tasks such as migration planning and business logic implementation.
In one test case, an agent completed a task in 18 hours, whereas a junior developer required over a week. Subsequent tasks, including Java and AngularJS plugin development, were completed in approximately 2 hours each. An architectural prototype for a Jasper-based report generator was developed in 10 hours.
The author explored multi-agent systems by assigning roles such as manager, architect, and developer to different agents. This approach was tested on a personal finance project, resulting in a functional application with a client, server, and database architecture.
The proposed development framework relies on isolated environments to manage tasks independently of human intervention. The architecture is designed to be adaptable for integration firms, allowing for deployment within a client's infrastructure or connection to external test benches.
