We are past the chatbot era. Products now delegate work—research runs, multi-file refactors, support resolutions—that unfolds over minutes or hours while the user is away. That changes what good UX means: the interface is no longer a sequence of screens to complete, but a contract for supervised autonomy. Plans, permissions, progress, rationale, and undo are the surfaces that decide whether people trust an agent with real stakes.
Agentic UX is among the highest-leverage design problems of this cycle. Model capability routinely outruns interface craft; teams ship autonomy before they ship visibility. When delegation fails, users do not blame the model alone—they blame the product that let work happen off-screen. Designing interfaces that align before action, show work in progress, and make recovery explicit is as foundational now as information architecture was for the web or touch patterns were for mobile.
This reference exists because that lifecycle lacked a shared vocabulary. Blog posts and one-off screenshots do not help teams audit a full run from goal to undo. Agentic UX names eleven patterns—before, during, and after action—with production examples, limitations, and Do/Avoid guidance so designers and builders can spec inspectable agent experiences instead of improvising every release.
How this relates to other references
Shape of AI catalogs AI UX broadly (prompting, media, trust, identity). Agentic UX focuses on autonomous agents acting on your behalf, organized by lifecycle phase so teams can audit a full run.
IF's catalogue centers policy, consent, and automated decisions in regulated contexts. Agentic UX complements that with run-time supervision mechanics—plans, ledgers, consent gates, undo—in developer and productivity tools.
Victor Yocco's Smashing Magazine article names six core lifecycle patterns for agentic AI. Agentic UX extends that set with additional pre-, in-, and post-action patterns, production screenshots, interactive mockups, and agent-readable specs.
Google PAIR's Guidebook covers human-centered AI broadly—when to use ML, trust calibration, feedback, and error recovery—via principles and patterns on hypothetical apps. Agentic UX narrows to long-running agents: run-time lifecycle surfaces (before / during / after a supervised run) with production screenshots. Use PAIR for foundations and automation tradeoffs; use Agentic UX to spec inspectable runs in Cursor, Replit, OpenAI Deep Research, and Perplexity.
See also the full Crosswalk mapping Agentic UX patterns to related work.
How to use this site
Read the framework article for the thesis, then follow the Guide to audit an agent flow in your product and turn gaps into pattern specs.