Building with AI · Agent UX patterns

Goal mode and activity streams

You can describe the goal-mode interaction pattern (state a goal, get a stream), how it differs from chat, and how an orchestrator's view scales it to multiple concurrent agents.

Chat is the wrong primitive once an agent runs for tens of minutes or hours. The chat metaphor — turn, response, turn — assumes the user is engaged moment-to-moment. Goal mode replaces that with a different shape: the user states an outcome, the agent runs against it, and an activity stream is the record. The user can leave, return, ask questions, redirect. The chat box is one input mechanism among several, not the spine.

Claude Code 2.1 shipped /goal as the explicit entry point for this. Replit Agent 3's 200-minute autonomous runs use the same shape. The 2026 Agentic Coding Trends Report frames it as the shift from "AI in the editor" to "orchestrator + specialized sub-agents on isolated git worktrees" — and the orchestrator's view is a goal stream, not a chat scroll.

Three things differ from chat:

Chapter contains 3 lessons.