CopilotKit maps three working patterns for letting agents control UI

CopilotKit's guide splits generative UI into three concrete tiers: static (agents pick from pre-built React components via useFrontendTool), declarative (agents emit JSON specs like A2UI or Open-JSON-UI that the frontend renders), and open-ended (agents hand over a full UI surface, as MCP Apps does, with the frontend just hosting it). The framework itself, AG-UI, isn't a UI spec but an event protocol coordinating tool lifecycles and state underneath all three. What makes this useful over the usual generative-UI hand-waving is that each tier comes with working code and an explicit control-versus-freedom tradeoff, so a team can pick static for safety-critical forms and open-ended only where embedding risk is acceptable.