Agent UX is the design discipline for products where a human user delegates meaningful work to an AI agent. It covers three surfaces: how the user tells the agent what to do (prompt surface), what the user sees while the agent works (progress surface), and how the user reviews or edits the agent's output (output surface). Every AI-powered product ships some version of all three, whether the designer thought about them or not.

How is designing for AI agents different from traditional UX?

Traditional UX assumes the user does the work and the interface enables it. Agent UX assumes the agent does the work and the interface manages the delegation. Different problems become primary: how to phrase requests to a probabilistic system, how to show progress on tasks with variable duration, how to review output when the user could not have produced it manually, how to build trust when the agent will occasionally be wrong.

What are the main design surfaces of an agent product?

Three: the prompt surface (input phrasing, examples, templates, disambiguation), the progress surface (streaming vs batch reveal, sub-task visibility, interruptibility, latency framing), and the output surface (accept/reject/edit affordances, diff views, regeneration paths, confidence indicators). Most agent products design one surface well and neglect the other two.

Are chat interfaces the best UX for AI agents?

Chat is the training-wheel UX for the agent era, not the destination. It works when the task is conversational and open-ended. It fails when the task has structure, when the user needs to review a specific piece of output, or when the agent's work should happen in the background of another workflow. Expect chat-first products to eventually surface parts of the work in more purpose-built UIs.

What are common mistakes in agent UX?

Five recurring ones: confidently wrong output with no error signal, no observable state during long agent runs, no way to interrupt an agent mid-task, showing the entire chain-of-thought when the user wants a summary, and hiding the chain-of-thought when the user wants to debug. Almost every AI product ships with at least two of these.

Agent UX in 2026: the three surfaces of designing products where AI does part of the work

2026-08-22

Every product is quietly becoming an agent product. The user still opens the app, but a large chunk of the actual work is being done by an AI agent the user doesn't fully control. Notion has agents drafting pages. Linear has agents grooming backlog. Gmail has agents summarizing threads. Cursor has agents writing code. And thousands of new products are shipping this year where the agent is the product. The old UX playbook doesn't cover this. Form fields, buttons, tables, modals: those are for products where the user does the work and the UI enables it. When a probabilistic agent is doing significant work on the user's behalf, different questions become primary. How do you phrase a request when the input matters more than the button placement? How do you show progress on a task with variable duration? How do you review output when you couldn't have produced it manually? Agent UX is the design discipline that answers those questions. This piece is the map: what the category actually is, the three surfaces every agent product ships, the failure modes almost every product falls into, and where design for this era is going next. ## Where "agent UX" comes from Product designers have been sneaking up on agent UX for two decades without calling it that. **Form wizards were the ancestor.** Multi-step guided flows (TurboTax, insurance quote engines, checkout wizards) are the earliest UI pattern where the user was answering questions and an intelligent system (rules-driven, not AI)

Agent UX in 2026: the three surfaces of designing products where AI does part of the work

Related posts