57% of enterprises have watched AI agents be confidently wrong. The fix is an agentic context layer, but who has one?