“It’s the Architecture, Stupid” — Why Prompt Engineering Won’t Fix Agents

This is a strong framing. I especially like the distinction between the model and the architecture around it. Treating the model as the execution engine, rather than the whole agent, makes the problem much clearer. The shift from Prompt → Model → Output to Skill → Execution → Structured Output also seems important, because it turns vague reasoning into reusable operational units. One small point I would add: clarification should also be part of the architecture, not just a prompt instruction. If a skill does not have enough required input, the agent should not silently infer the missing context. It should enter a clarification state and ask before execution. In that sense, “ask if unsure” should be a runtime behavior, not just wording inside a prompt.