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

Completely agree. Clarification as a runtime state, not a prompt instruction, is a much more reliable design. If a required input is missing, the skill should block and ask rather than silently fill the gap with inference. That’s exactly the kind of behavior that needs to live in the execution layer, not in the model’s judgment. We have partial coverage: agent.input.collect handles structured field collection with explicit missing_fields tracking, and agent.goal.interpret outputs a requires_clarification flag when open questions exist before planning can proceed.

What we don’t have yet is a first-class clarification state enforced at the execution layer itself, meaning the runtime doesn’t automatically block and enter a clarification loop when required inputs are absent. That decision still falls to the orchestrator consuming those contracts.

Your point is a good design target: clarification should be a runtime primitive, not an optional output field. Adding it to the roadmap.