10 Agentic UX Patterns and Real-Life Scenarios

An agentic UX pattern is a repeatable piece of interface for handing work to an AI agent and staying in charge of the result. Our agentic interaction patterns reference describes each one in the abstract: what it is, when it fits, and what it costs. This post takes ten of them and puts a shipping product and a scenario beside each.

Eight are about where the agent sits in a product. The last two are about how it behaves once it is working. Every product example comes from the vendor's own documentation, checked on 11 October 2026.

1. Entry point: one way in, reachable from everywhere

The entry point is the control that starts an agent interaction: a persistent button, a command palette, or a prompt field in the header.

In a real product. Shopify's assistant, Sidekick, is not tied to one screen. Shopify's help centre says a merchant can chat with it from "any page in your Shopify admin on desktop or mobile", and that Sidekick "presents changes for your review before applying them".

Scenario. A shop owner is looking at an order that shipped to the wrong address. The same button she used yesterday on the discounts page is there on the order page, and it already knows which order she means.

The cost is that a floating button is the most ignored element on most pages. An entry point needs a reason to be pressed, and "Ask AI" does not supply one.

2. Inline chat: the conversation sits on the thing it is about

Inline chat is a conversation embedded in the page, next to the content it concerns.

In a real product. Visual Studio Code opens inline chat inside the editor with Cmd+I, or Ctrl+I on Windows and Linux. Per the VS Code documentation, "VS Code shows a diff with the code suggestion inline in the editor", and the developer presses Keep or Undo.

Scenario. An analyst selects a paragraph in a quarterly report and asks for it to be half as long. The shorter version appears in place, marked as a change, and one click keeps it.

Inline chat suits short, local requests. A conversation that outgrows its box leaves the person scrolling inside a scroll.

3. Sidebar chat: a panel that stays open while the work continues

Sidebar chat is a docked panel beside the main content.

In a real product. Gemini in Google Docs opens from an "Ask Gemini" button at the top right and works in a side panel, where it can summarise the open document or rewrite a section. Google's help page adds a warning: the panel's history is lost on refresh, so "to avoid the loss of your history, insert generated output in the document".

Scenario. A lawyer reviews a lease on the left and asks questions about it on the right. When the agent drafts a replacement clause, she inserts it into the lease, because the panel will be empty tomorrow.

That warning is the pattern's main lesson. If the agent produced something worth keeping, it belongs in the document and not only in the panel. The panel also takes a third of the width whether or not anyone is using it.

4. Full-screen chat: the conversation is the product

In full-screen chat there is no surrounding application. The prompt field and the transcript are the whole interface.

In a real product. Claude, ChatGPT, and Perplexity all work this way. The detail worth copying is that a conversation gets an address. Anthropic's support documentation describes sharing as creating "snapshots of your conversations to share via direct link", and says "anyone with the link can view the chat snapshot".

Scenario. A founder spends an hour working through a pricing model with an agent, then sends her co-founder a link to the result. He reads the same conversation she had, without a screenshot or a pasted transcript.

Without stable links, nothing a full-screen agent produces can be cited, bookmarked, or returned to. The other cost is discovery: everything the product can do has to be findable from an empty field.

5. Contextual AI actions: named operations on a specific object

A contextual AI action is a labelled operation offered on one item, such as "summarise this" or "explain this error".

In a real product. Slack puts a Summarize button at the top of a thread. Slack's guide to its AI features says the summary appears when it is ready, and a "More details" control shows the messages it drew on.

Scenario. A project manager returns from a week away to a 140-message thread about a delayed launch. One button gives her the decision, who made it, and links to the three messages that matter.

This is the pattern to reach for before chat. A named action says what will happen, which a blank prompt cannot. It is also the easiest pattern to expose to other software: an action with a name, an input, and a result is a tool.

6. Feed states: show what the agent is doing while it does it

Feed states are what the interface shows during the work: the step in progress, the tool that was called, the result that came back.

In a real product. GitHub's Copilot coding agent works in the background and pushes commits to a draft pull request. In GitHub's launch post, the company says "you'll see the agent's reasoning and validation steps in the session logs".

Scenario. A travel agent is booking a multi-city trip. The screen reads "Checking seat availability on the 09:40", then "Fare changed, re-pricing". The traveller sees the price move before the agent commits to it, and stops the task.

Narrating every internal step buries the one that mattered. Show what was done to the outside world, and keep the reasoning one click away.

7. Citations: the claim and its source, attached

A citation ties a statement to the passage it came from, closely enough that a reader can check one against the other.

In a real product. Anthropic's API has a citations feature that, in its documentation, is described as returning "the exact passages that support each claim". The passage is the unit, and the file is not.

Scenario. A compliance officer asks an agent whether a supplier contract allows subcontracting. The answer says no and shows the sentence in clause 14.2 that it relies on. She reads one sentence and moves on.

A citation that points at a 40-page PDF is decoration. The work in this pattern is finding the right passage, and it is the only cheap way for a person to verify an agent without redoing the job.

8. Disclosure and limits: say what the agent is and what it cannot promise

Disclosure tells people they are dealing with an AI system, where it is unsure, what it may not do, and how to reach a person.

In real life. Two examples, and neither is a product feature. Article 50 of the EU AI Act requires providers to design AI systems so that people "are informed that they are interacting with an AI system", and it has applied since 2 August 2026.

The second is a ruling. In Moffatt v. Air Canada, decided on 14 February 2024, British Columbia's Civil Resolution Tribunal ordered the airline to pay a passenger $650.88 in damages after its website chatbot wrongly told him he could claim a bereavement fare after travelling. Air Canada argued the chatbot was responsible for its own statements. The tribunal called that a remarkable submission and held the airline responsible for everything on its website.

Scenario. An insurance agent answers a coverage question, then adds: "This is a general answer. I can't confirm your policy's exclusions. Here is the clause, and here is how to reach a claims handler."

A blanket "AI can make mistakes" in the footer protects nobody, and the Air Canada case suggests it would not have helped. Specific uncertainty on a specific claim is the version that works.

9. Invisible handoff: report the work, not the worker

When one agent delegates to another, an invisible handoff keeps a single conversation going. The person hears what is being done in their own terms and never has to learn the architecture.

In a real product. Anthropic's research feature is several agents. In its engineering write-up of 13 June 2025, a lead agent "analyzes it, develops a strategy, and spawns subagents to explore different aspects simultaneously", and a separate agent places the citations. The user receives one report.

Scenario. A customer asks a bank's assistant why a payment bounced. Behind the reply, one agent reads the transaction log and another checks fraud rules. The customer sees "Checking your recent payments", then one answer, and is never told she has been transferred.

Anthropic's post is candid about the price: multi-agent systems "use about 15× more tokens than chats". The handoff being invisible to the user does not make it free. Our page on entry points and orchestration covers what the front agent owes the user when a specialist fails.

10. Progressive automation: start with suggestions, earn the rest

Progressive automation offers several levels of independence, so a person can begin with suggestions and delegate more as trust grows.

In a real product. Linear's Triage Intelligence lets each team choose whether suggestions for a property type "appear, are hidden, or are auto-applied". Labels can apply themselves while assignees still wait for a person. Claude Code does the same for a whole session: its documentation lists six permission modes, from asking before every edit to running everything, and a developer switches between them with Shift+Tab.

Scenario. An accounts team lets an agent suggest expense categories for a month. Once nine in ten suggestions go through unedited, they turn on auto-apply for travel and meals, and leave anything over $1,000 on manual.

Set the level per type of task. A single global switch forces people to choose for their riskiest job, and then the agent asks permission for everything.

Ten more, for how the agent behaves

These ten are mostly about placement. The patterns page now also carries ten behaviour patterns, from intent preview to editable memory, each with its own product example. Six of those names come from Victor Yocco's Smashing Magazine article and from Eleken's six agentic UX patterns, and four are ours.

One caution about every example in this post. Each is what the vendor documents, and we have not measured how any of them performs with real users.