<span>Procedural Content Generation (PCG) provides a means of automatically&nbsp;</span><span>generating diverse, complex functional environments for video game playing&nbsp;</span><span>agents. While various optimization approaches have been used to generate&nbsp;</span><span>content meeting computable functional constraints, more open-ended human priors&nbsp;</span><span>and directives are difficult to bake in to such methods. We propose Agentic&nbsp;</span><span>Procedural Content Generation, a paradigm in which tool-calling large language&nbsp;</span><span>models generate video game levels that are both functional and steerable via&nbsp;</span><span>natural language. Equipped with local brushes, generative algorithms, and&nbsp;</span><span>evaluation functions, agents iteratively refine content toward specified design&nbsp;</span><span>objectives. We show that the framework applies to both static level design&nbsp;</span><span>tasks and environments with dynamic gameplay mechanics, while satisfying</span> <br><span>intended functional constraints. Beyond such constraints, the same framework&nbsp;</span><span>can respond to open-ended gameplay prompts and produce levels with stronger&nbsp;</span><span>structural and stylistic coherence than purely metric-driven search, suggesting&nbsp;</span><span>that Agentic PCG is a framework for balancing functional constraint&nbsp;</span><span>satisfaction with free-form design control.</span>