[D] Predicting total cost of agentic LLM workflows - is there a research gap around output token count and chain depth estimation?

About this article

Working on a practical problem that I think has an interesting ML angle. In agentic LLM workflows (tool use, multi-step reasoning, ReAct-style loops), the total cost of a task is a function of: - Output token count per step (unpredictable before generation) - Number of chained calls / loop depth (emergent, depends on intermediate results) - Context growth across the session (each step adds to the next call's input) - Cache hit rates (variable, affects pricing by up to 90%) Input token count i...