AI Agents: Interaction

Agent UX is maturing into its own discipline. The core constraint is "conversation-native" design — UIs optimized for chat width, scroll, and inline rendering rather than dashboard layouts (Tool UI rendering JSON outputs as inline, narrated, referenceable surfaces) — and a new customization category of skills that shape how the agent communicates, not just what it does, because "cognitive debt" erodes the productivity gain when output is hard to parse. Human-AI interaction design is consolidating around shared pattern libraries (the AI Interaction Atlas) rather than per-product reinvention, with the everyday wedge being leverage on managerial ritual: AI compressing a Friday review to ~12 minutes and 1:1 prep to 5, calendar-aware proactive scheduling, and Wargame.esq exposing two agents' real-time reasoning as they negotiate a contract point-by-point — transparency into AI decision-making as a first-class UX goal.

The deepest framing is intelligence as a social process: frontier reasoning models spontaneously generate "societies of thought" (internal multi-agent debates that causally drive accuracy, discovered through RL alone), every prior intelligence explosion was a new socially aggregated unit of cognition rather than an individual upgrade, and the path to more powerful AI runs through composing richer human-AI social systems — "centaurs" in shifting configurations, intelligence growing like a city, not a single colossal oracle. Adversarial prompting is also maturing as a first-class interaction discipline: Marc Andreessen's expert-persona prompt (lead with the strongest counterargument, never capitulate without new evidence, use explicit confidence levels, don't anchor on user estimates) operationalizes anti-sycophancy; Roughdraft.md provides a local-first Markdown review surface for human-AI collaborative editing; and Fable's "read the relevant academic literature — then think adversarially" combines research grounding with systematic challenge as a pre-build ritual.