Agent UXResearch report

Abstract floating conversational interface panels connected by glowing threads

The design patterns of the AI-native interface

Q: Is "Agent UX" a real discipline or just a trend term?

It is real, with a caveat. The trust, transparency and control principles have a peer-reviewed backbone: Microsoft’s 18 Guidelines for Human-AI Interaction (CHI 2019), validated with 49 practitioners, plus the HAX Toolkit pattern library. The specific interface patterns layered on top, citation pills, token UX, video-call avatars, are widely practiced but not yet authoritatively named. The spine is solid; the surface is emerging convention.

Q: What should I call it: Agent UX, Agentic UX, AI UX or AX?

For search and conversation, "Agent UX" and "AI UX" are the dominant practitioner terms. For authority, anchor on "Human-AI Interaction", the academic lineage from the Microsoft / CHI work. Avoid leaning on "Agent Experience (AX)" as an established discipline; it is largely one vendor’s framing and did not survive independent verification.

Q: Why did these patterns not exist two years ago?

Because the technology did not demand them. Conversational, non-deterministic, tool-using, metered agents create problems deterministic software never had: uncertainty you must communicate, actions you must authorise, costs that accrue per turn, and a personality users react to. Each pattern is a response to one of those new realities.

Agents are conversational, uncertain, tool-using and metered. That created a new set of UX problems, and a new set of patterns to solve them. Here is the reference, with each pattern graded by how well it is evidenced, and live demos.

Explore the patterns

Published 18 June 2026
7 deep dives
Live demos
Graded by evidence

(01) The series

The deep dives

Each pattern gets its own piece, with a live demo you can feel. More are landing in this series.

Calibrated confidenceVerified spineSurface how sure the agent is, in a way that prevents overreliance rather than manufacturing it.Tool-call transparency and write-confirmVerified spineShow what the agent is doing, and confirm before it changes anything in the real world.Inline citations and source groundingObserved conventionAttach the source to the claim, in the answer, as a pill or footnote the user can open.Token and cost-transparency UXObserved conventionShow the balance and what each turn costs, because metered AI creates real anxiety.The video-call frame for voice agentsObserved conventionBorrow the video-call layout so talking to an AI feels familiar: agent as the main character, you as the small self-view.Naming the agentObserved conventionA named agent with a defined voice is a product. A generic AI assistant is a feature nobody remembers.Artifacts versus the chat streamContested frontierSome outputs cannot live in a scrolling transcript. They need their own surface.

(02) The context

Why a new discipline appeared

For thirty years we designed for deterministic software. You clicked a button, the same thing happened every time, and the interface’s job was to make the available actions obvious.

Agents broke that contract. They are conversational, so the chat input became the primary surface. They are uncertain, so you have to communicate how sure they are. They take actions in the real world, so you have to show what they are doing and ask before they do it. They cost money per turn, so usage became something users watch with anxiety. None of these were interface problems two years ago.

(03) The verdict

Not all of these patterns are equal. Some sit on a peer-reviewed foundation. Some are widely practiced but barely documented. A few are vendor hype dressed up as a paradigm. This report separates them honestly, because building on the wrong one is how you lose a user’s trust.

Verified spine

patterns on a peer-reviewed or research-backed spine

Observed convention

patterns widely practiced but barely documented, the white space

Contested frontier

real products oversold as a settled paradigm

(04) The taxonomy

The pattern taxonomy

Eight patterns, honestly graded. Select one to see what it is and how well evidenced it is.

Calibrated confidence

Verified spine

Surface how sure the agent is, in a way that prevents overreliance rather than manufacturing it.

Every agent answer carries uncertainty. The naive fix is to show the reasoning trace and let the user judge. The research says that backfires: a fluent explanation reads as a competence signal and makes people trust a wrong answer more, not less.

The pattern that works is calibrated confidence: a small, honest signal of certainty, the specific part the agent is unsure about, and a one-tap way to verify. Confidence theatre is the anti-pattern.

Evidence

Verified. Microsoft Design (2025) requires certainty and reasoning to be accessible to avoid overreliance. ACM 2026 research documents the transparency backfire where fluent explanations increase misplaced trust.

Read the full pattern

(05) The timeline

How we got here: 2022 to 2026

Each pattern is a response to something that changed in the technology.

Chat becomes the surface
ChatGPT launches (Nov 2022). The conversation, not the form, becomes the primary interface. The persistent input fixed at the bottom of the screen arrives as the default.
The first agent affordances
Streaming responses, stop generating, regenerate and thumbs up/down feedback become standard.
Bing Chat and Perplexity push inline citations into the mainstream. Grounding becomes an expectation.
The first plugins and tool calls appear. Users start needing to see what the model is doing on their behalf.
Output leaves the chat
Claude Artifacts (Jun 2024) and ChatGPT Canvas (Oct 2024) move documents, code and apps out of the transcript into dedicated surfaces.
The Model Context Protocol (Anthropic, late 2024) standardises tool calling, which forces a consistent connect, authorise and act flow.
Agents act, and ask first
Agentic modes, computer use and plan-style review normalise the read-freely / write-confirm pattern.
Voice and video agents adopt the familiar call layout. Realistic and abstract avatars compete.
Microsoft Design publishes explicit UX guidance for agents: actions must be visible and controllable.
Trust and cost get serious
Token and cost transparency become a felt UX requirement as metered usage spreads.
The transparency backfire debate lands: ACM research shows naive reasoning displays can increase overreliance, raising the bar to calibrated confidence.
The EU AI Act transparency obligations (Aug 2026) turn several of these patterns from good practice into expectation.

(06) The sources

The foundations

The field has a real, citable backbone. These are the sources the durable patterns rest on.

(07) Questions

Frequently asked questions

Is Agent UX a real discipline or just a trend term?

It is real, with an important caveat. The trust, transparency and control principles have a peer-reviewed backbone: Microsoft’s 18 Guidelines for Human-AI Interaction (CHI 2019), validated with 49 practitioners, plus the HAX Toolkit pattern library. The specific interface patterns layered on top (citation pills, token UX, video-call avatars) are widely practiced but not yet authoritatively named. So the spine is solid; the surface is emerging convention.

What should I call it: Agent UX, Agentic UX, AI UX or AX?

For search and conversation, Agent UX and AI UX are the dominant practitioner terms. For authority, anchor on Human-AI Interaction, the academic lineage from the Microsoft / CHI work. Avoid leaning on Agent Experience (AX) as if it were an established discipline; it is largely one vendor’s framing and did not survive independent verification.

Which of these patterns are safe to build on, and which are hype?

Safe: tool-call transparency, write-confirm, calibrated confidence and inline citations, all backed by guidance or research. Real but oversold as a paradigm: generative UI and the new agent-UI protocols (MCP-UI, AG-UI, A2UI). Use the products, do not bet your architecture on the manifesto. Genuinely under-documented and therefore an opportunity: token-cost UX and the video-call frame for voice agents.

Why did these patterns not exist two years ago?

Because the underlying technology did not demand them. Conversational, non-deterministic, tool-using, metered agents create problems that deterministic software never had: uncertainty you must communicate, actions you must authorise, costs that accrue per turn, and a personality users react to. Each pattern in this series is a response to one of those new realities.

Does p0stman actually build these?

Yes. These are not slides. p0stman ships named voice and video agents, inline citations, read-freely / write-confirm tool flows and metered usage across its own products and client work. The live demos in this series run on the same components we put into production.

Take it with you

Don’t just read this. Put it to work.

The whole series is distilled into one Markdown file: every pattern, the do and don’t rules, and how well each is evidenced. Download it into your project, or paste the link into any chat with your agent and tell it to improve your agent UX. It’s free, no sign-up, no attribution required.

Paste this into your agent

Use these Agent UX principles to review and improve our agent's interface: https://p0stman.com/agent-ux/agent-ux-principles.md

Download the .md

We build agents, not slideware

The clearest, most under-served pattern in the whole field is also the one most likely to break trust if you get it wrong. We wrote the definitive piece on it, with a live demo of confidence theatre versus calibrated confidence, so you can feel the difference.

Talk to Paul about your agent