Plan–Code–Execute: Designing Agents That Create Their Own Tools

A plan–code–execute agentic architecture is proposed as an alternative to using predefined toolsets for AI agents. This system dynamically generates its own tools as executable Python scripts to solve complex problems as they arise. The architecture is composed of five distinct agents: an Analyst for data discovery, a Planner to decompose tasks, a Coder to write the scripts, an Executor to run them, and a Reporter to summarize the findings. This approach is demonstrated by building an Explainability Agent to analyze the results of a Graph Neural Network model for demand forecasting. The goal is to create more flexible and grounded AI systems that can adapt to new tasks without pre-built tools.