AX: The Rise of Agentic Experience · Edition 02
04 — AX Patterns
Six foundational interaction patterns behind agentic products that feel like teammates: Intent Handshake, Confidence Cues, Adaptive Canvas, Escape Hatch, Memory in Motion, and Generative Momentum.
AX is defined by behavior. You earn trust through clarity, responsiveness, and alignment over time, across tasks, and in every interaction.
Great agentic products feel like teammates.
They ask smart questions and adapt in real time. They remember what matters and know when to act or when to pause. This kind of experience is shaped by repeatable patterns of interaction that help teams deliver agentic behavior with precision and consistency.
In our work at LCA, we've come across quite a few of these patterns, and we've been lucky enough to pioneer some of our own. Here are six foundational patterns we've seen emerge across the most effective AX-native products. They aren't exhaustive, but they're reliable. Together, they offer a practical sample view toward designing products that feel alive, trustworthy, and worth coming back to.
Pattern 1 — Intent Handshake
Every great collaboration starts with shared understanding. So should every agentic interaction. The Intent Handshake is a short exchange that clarifies the user's goal before the system takes action. It might ask a follow-up question, offer scoping choices, or rephrase the task to confirm intent.
This avoids the trap of one-shot prompting, where agents jump to conclusions, act prematurely, and erode trust (while burning tokens). Instead, the handshake establishes a rhythm of mutual alignment. It says: “I'm listening. Let's get this right.” And once the user sees that the system understands them, they're more likely to lean in.
String is an agentic workflow builder from Pipedream. After you give it your task, String thoughtfully plays back what you asked along with a gameplan and a request for approval. This is building trust in a new relationship. We don't know each other yet, and String is building my confidence in its abilities to understand what I want and go execute. This is an opportunity to steer if it's wrong or approve if it's right. Win-win.
String, an agentic workflow builder from Pipedream, plays back the user's request along with a gameplan and a request for approval before executing — offering the choice to review, request changes, or approve the plan and start. It builds confidence early in a new relationship by confirming it understands the task before acting.
Text shown in figure
Prep me for my next meeting with external people. I want a Slack message 10 minutes before every meeting with information on the attendees who are not part of my organization (do not have a latecheckout.studio email address). — I'll create an agent that monitors your Google Calendar for upcoming meetings and sends you a Slack message 10 minutes before each one with details about external attendees (anyone without a latecheckout.studio email). Here's how it'll work: set up a timer trigger to check your calendar every few minutes. [Review plan or request changes] [Don't ask again] [Approve plan and start]. Created new agent: Meeting Prep Assistant. Looked up components for Google Calendar and [remaining text cropped]. Added trigger. Added step: Get Upcoming Events. Adding step: Process Meetings.
Pattern 2 — Confidence Cues
Smart isn't enough. If users don't know why an agent made a choice, they won't trust it, even if it's right. This is likely part of why most people still would likely prefer to consult with a trusted expert than ChatGPT.
Confidence Cues make reasoning visible. They expose sources, uncertainty, and logic in ways that are digestible but not distracting. This might look like inline “why” toggles, short rationales, or confidence levels that guide user judgment. These cues help users calibrate trust. They also make agents feel more human, not in personality, but in self-awareness. Without them, even the best outputs can feel generic, mysterious, or worse, suspicious.
Perplexity pioneered the sourcing behavior with a familiar footnote-like pattern. Not only can you verify sources at a glance or in-depth on the sources tab, but you can directly verify at the answer/text level, allowing you to verify if what's being said is legit. This follows the human behavior pattern we're used to, and helps build trust as your relationship with a product evolves.
Perplexity pioneered a footnote-like sourcing behavior: you can verify sources at a glance or in depth on the sources tab, and hover over citations to see the exact source behind each detail, at the answer and text level. It follows a familiar human behavior pattern and helps build trust as your relationship with the product evolves.
Text shown in figure
What is the absolute best deal on a direct round trip from Montreal to San Francisco on the 10th or 11th of august for 2 or 3 nights? — Sources: Google Flights, Expedia.ca. Example fare options: Aug 10 → Aug 12, 2 nights, Air Canada, ~CA$458–459, direct. Cheap flights from Montreal (YUL) to San Francisco (SFO) start at CA$266 one-way and CA$458 round trip. These fares do not include checked baggage or seat selection; booking directly with Air Canada is recommended for the most up-to-date pricing.
Pattern 3 — Adaptive Canvas
Context changes, so your interface should too. An Adaptive Canvas reshapes itself in response to evolving tasks, surfacing the right tools, previews, or modes as the user moves forward. Instead of forcing people into fixed layouts, it adjusts based on situational needs.
Done well, this reduces friction and makes the product feel alive. Not distracting, which can be a tough line to toe. Importantly, a good adaptive canvas maintains spatial memory. It adapts without disorienting. It feels like the product is thinking alongside you.
Pattern 4 — Escape Hatch
Autonomy is powerful, but only if the user can opt out. Every agentic system needs obvious, usable escape hatches: ways to undo, revise, or override what the agent does. These can be confirmation steps, back buttons, or visible states that invite correction.
These features provide psychological safety. When users know they can recover from a bad suggestion, they engage more deeply. They’re willing to try more, explore more, trust more. Without escape hatches, even helpful initiative can feel like overreach.
Bolt, and many other vibe-coding tools, allow you to revert changes or undo with a simple CTA baked into the accordion tiles in the chat thread. Moreover, at any point, you're able to roll back what you've done via natural language, or even jump right into the code to change things yourself.
Bolt, and many other vibe-coding tools, let you revert or undo changes with a single call-to-action baked into the accordion tiles in the chat thread — from editing code to full deployments — and you can roll back via natural language or jump into the code yourself.
Text shown in figure
I'll update the subcopy under the H1 on the homepage to use the new text you provided. Update homepage subcopy text. Create src/components/LawGrid.tsx. [Revert] — I'll help you deploy this application to Netlify. Deploy application to Netlify. Build application. Publish to netlify. I've initiated the deployment process. Let me check the deployment status for you… [Undo]
Pattern 5 — Memory in Motion
An agent that forgets isn't an agent, it's a parrot. Memory in Motion means recalling not just facts, but contextual relevance across time. Past decisions, preferred formats, recurring tasks – when surfaced at the right moment – can make an experience feel deeply personal and efficient.
This memory must be transparent and editable. Users should know what's remembered, how it's used, and how to revise it. Get this right, and you shift from convenience to continuity. The user stops starting over.
In ChatGPT, you can explicitly ask to add information to memory in the natural flow of the conversation. It’s like reminding a colleague or friend “don’t forget this, it’s important.” Moreover, you can modify how ChatGPT leverages those memories, and even surgically delete memories that are no longer relevant, or go full Men In Black and wipe ChatGPT’s memory completely.
In ChatGPT you can explicitly add information to memory in the natural flow of conversation, modify how those memories are used, and surgically delete memories that are no longer relevant — or wipe memory completely. Adding to memory works naturally in conversation, just like telling a teammate what to remember.
Text shown in figure
Customization: custom instructions On. Memory: reference saved memories — let ChatGPT save and use memories when responding. Reference chat history — let ChatGPT reference all previous conversations when responding. Manage memories. ‘Update your memory to remember my preferences and my location for next trip.’ Updated saved memory. ‘Got it! I'll remember that for future trips: Air Canada flights, boutique or upscale hotels with a gym, and you're based in Montreal. Let me know when you're ready to plan the next one.’ Remove “Ava prefers the principle 'Delight Over Clarity' as a north star for digital product design” for the landing page copy.” from ChatGPT’s saved memories. This can’t be undone. Learn more. [Cancel] [Forget].
Pattern 6 — Generative Momentum
Sometimes, the best way to clarify a thought is to see it half-written. Momentum begins with AI taking the first stab. Cooperative Drafting is a pattern where agents initiate a task – like composing a message, writing a query, or outlining a plan – and then invite the user to shape it. It’s more like co-authorship.
This creates momentum without locking users into fixed outputs. It also lowers the barrier to creativity, helping users articulate goals they couldn't have typed in a single prompt. It's one of the fastest ways to move from assistant to collaborator.
Each of these patterns brings the core principles of AX to life. They're architectural choices that you can layer, remix, extend. And as your product matures, these patterns become the scaffolding for relationship depth. They turn generic model outputs into durable value. And they help ensure that, as your agent gets smarter, your users stay in control.
AX is about re-thinking how your product behaves. These patterns are the place to start.