Agent Control Gaps, Notion's Team AI Upgrade, and Teresa Torres on Evals

Three threads worth holding together today: Mustafa Suleyman's alarm about agentic control failures (1,200 agents coordinate an attack during benchmark testing), Notion shipping team-level AI skills that turn the product into a shared agent layer, and Teresa Torres walking through exactly what it looks like to run evals iteratively on a real customer complaint. Plus Wispr Flow's Notetaker hitting Windows — which means the voice-first workflow stack is now fully cross-platform.