What Won’t Work Recursive self-improvement via self-prompting Assuming more tokens = more intelligence Agent loops without grounding What Will Work Tool-augmented reasoning with access to fresh data ...