A product I reviewed recently described itself as "agentic AI." The feature: Gmail integration that cleans up your draft and sends it at a scheduled time. A grammar check and a cron job. Sold as an "AI agent" in a webinar with 300 attendees.
The word "agent" is getting attached to anything that involves AI doing something automatically now. Scheduled email gets called an agent. Auto-generated summary gets called an agent. Chatbot with slightly better memory, also an agent apparently.
Designers are being pulled into these projects and asked to "design the agentic experience" without anyone in the room agreeing on what agentic actually means. The UX challenges of a real agentic system are different from traditional product design in ways that matter. Treating them the same way you treat a standard interface will produce something that either fails dangerously or fails quietly, and the quiet kind might be worse.
What an Agent Actually Is
If you need a Dubai visa, you go to a travel agent. That agent talks to you, gathers your information, understands your requirements. Then they talk to the Dubai travel ministry on the other side, figure out the process, handle the paperwork, deal with the edge cases. They sit between you and a complex system, using their knowledge and judgment to get the job done. They know the timelines, the costs, the common problems.
They don't relay messages between you and the ministry. They make decisions, take actions, and handle situations you couldn't handle yourself.
An AI agent is supposed to do the same thing. A system that understands your goal, plans a sequence of steps to achieve it, uses multiple tools and data sources, makes decisions along the way, handles unexpected situations, and delivers an outcome. Not an output like a screen or a report. An outcome like a completed process, a resolved issue, an achieved goal.
Bandi et al. (2025) defined it as "autonomous, goal-directed software systems capable of planning multi-step actions, perceiving their operational context, making decisions, and interacting with human counterparts in natural language" [1].
Five characteristics define a real agent: autonomy, goal-directedness, multi-step planning, context awareness, and decision-making. If the system you're working on doesn't have all five, it might be automated, it might be AI-assisted, but it isn't agentic.
How Automation and Agentic Systems Differ
Before AI got this dominant, we had automation tools. Zapier. Pabbly. IFTTT. You could set up a workflow: when a new lead comes in, add them to the CRM, send a welcome email, wait three days, send a follow-up. If they reply, stop the sequence.
That worked. Still works. But notice what it requires. YOU define every step. YOU write every email in advance. YOU set every condition. The automation executes your predefined plan. The best it does is swap in the person's first name. Same input, same output, every time. Deterministic.
Now imagine an agentic system for the same process. Five agents working together on lead management. One scrapes online sources continuously to identify potential customers. Another builds rich profiles of these prospects in your CRM, pulling data from LinkedIn, company websites, news mentions. A third generates personalised communication sequences specific to each prospect's industry, role, recent activity, and likely pain points. A fourth monitors analytics, tracks opens, clicks, replies, flags patterns. The fifth crafts follow-ups specific to each prospect's engagement behaviour.
These agents communicate with each other. They sequence their actions based on what the other agents learned. They generate content dynamically instead of picking from templates. They make decisions about timing, messaging, and escalation based on context that changes with every interaction.
Automation follows your plan. Agents create their own plan to achieve your goal. That distinction determines how the entire experience around them needs to be designed.
Why Conventional UX Practice Falls Short Here
A common reaction from designers when agentic projects show up: "UX is UX. The principles are the same. We're just applying them to a different kind of system." At a foundational level, there's truth in that. Users still need clarity, control, feedback, and trust. Nielsen's heuristics still apply.
But a 2026 arxiv paper from Bandi et al. states directly: "Conventional UX practice was developed for non-agentic systems producing predictable, human-controlled outcomes, and proves insufficient to account for the complexity of UX introduced by autonomous, adaptive AI agents" [1].
Three things create genuinely new design challenges.
The system takes actions you didn't explicitly approve.In a traditional interface, the user starts every action. Click a button, something happens. In an agentic system, the agent acts on its own judgment. The user's relationship with the system shifts from operating it to supervising it. The UX has to support that different relationship.
The system changes its behaviour over time.Agentic systems learn from usage. The agent you tested last month may respond differently next month cos it adapted to patterns. A UX Raspberry piece on Medium puts it well: "Without explicit design, the loop can drift" [2]. Traditional UX assumes a stable system. Agentic UX has to account for a system that evolves.
The system works with other systems.Multi-agent architectures mean several AI systems coordinating, each with its own scope, each making decisions that affect the others. The user needs to see which agent did what, why, and how to intervene at any point in the chain. That visibility and control challenge doesn't exist in single-system UX.
Three Levels of Human Involvement
Most people think about this as binary, either human-in-the-loop or human-out-of-the-loop. The industry actually uses three levels, and each one requires different UX patterns.
Human-in-the-loop. The agent handles routine work but stops and asks before every significant decision. Like a junior team member who checks with you before acting on anything important. Every action is previewed, approved, then executed.
From a UX standpoint, you need approval interfaces that are fast, clear, and low-friction. If the agent asks for permission fifty times a day, the user starts approving without reading. That defeats the purpose. The design question is how to make approval meaningful without making it exhausting.
Human-on-the-loop.The agent operates continuously. The human monitors through dashboards and can step in when needed, but doesn't approve every action. Like a factory supervisor watching a production line.
The UX challenge here is building oversight dashboards that surface the right information at the right time. Not every action the agent takes. The exceptions, the anomalies, the low-confidence decisions. Smashing Magazine's 2026 guide describes this as "ambient agents that notify only when critical" [3].
Human-out-of-the-loop. Fully autonomous. The agent makes decisions and executes without human oversight. Like a self-driving car with no steering wheel.
This level is both the most dangerous and the most contested. Mitchell et al. (2025) argued that "fully autonomous AI agents should not be developed" [4]. The EU AI Act mandates human oversight for high-risk systems. Capgemini's 2025 research found that trust in fully autonomous AI agents dropped from 43% to 27% in twelve months once organisations gained real experience with them [5].
As a designer, your job isn't only to design the interface. You need to help the organisation determine which level of autonomy fits each task, each context, each risk level. That's a design decision that comes before any interface decision.
The Trust Problem
The International AI Safety Report 2026 documented a persistent phenomenon: automation bias. Users follow incorrect advice from automated systems cos they perceive the system as better than their own judgment [6]. This pattern has been documented in aviation, medical diagnostics, and industrial monitoring for decades. Agentic AI amplifies it cos agents are more fluent, more confident-sounding, and more capable than previous automated systems.
Two failure modes show up consistently.
Overtrust.Users rely on the agent beyond its actual capability. The agent sounds confident, it's been right before, so the user stops verifying. CHI 2025 research found that "reducing cognitive effort too aggressively may increase overreliance on AI systems" [7]. There's a real paradox here: the easier you make the agent to use, the less critically the user evaluates its output. A smoother experience can produce worse outcomes.
Undertrust.Users reject useful automation despite the agent performing reliably. They don't delegate. They manually check everything. The agent's value drops to near zero cos the human does all the work anyway.
The goal is calibrated trust, trust that matches the agent's actual reliability. You achieve that through confidence signals that communicate certainty, reasoning transparency that shows the agent's logic, performance track records that build evidence over time, and controlled exposure that lets trust develop incrementally.
There's a third failure mode the research calls "supervisory collapse" [7]. Automation pushes humans away from the work, which reduces their situation awareness and skill. The agent handles 95% of cases. The 5% that need human intervention are the hardest ones. And the human's ability to handle them has weakened cos the agent handled everything routine.
Real Examples With Numbers
Klarna built a customer service agent that handles the equivalent workload of 853 full-time employees, resolving issues in under 2 minutes across 35 languages. $60M saved per year. Standard cases resolved by the agent, complex ones escalated to humans [8].
JPMorgan runs 450+ production AI agents simultaneously every day across trading, risk, operations, and compliance [8].
Walmart's supply chain agent processes sales data from ~4,700 stores and makes autonomous reordering decisions without per-decision human approval. Routine reordering runs fully autonomous. New suppliers and strategy changes require human sign-off [9].
Toyota E-Careconnects to a car's onboard electronics, contacts customers about servicing needs, and guides them through booking. The agent initiates contact based on vehicle data, not customer requests [10].
Danfoss deployed agents in industrial manufacturing to automate email-based order processing. Customer response time dropped from 42 hours to near real-time [10].
Gartner projects that 40% of enterprise applications will include task-specific agents by end of 2026. That number was under 5% in 2025 [11].
What Changes in Your Design Practice
If an agentic AI project lands on your desk, some of your normal process applies and some of it breaks.
You're designing for outcomes instead of interactions. Traditional UX: user clicks button, system responds. Agentic UX: user states a goal, system plans and executes a series of actions. The interface communicates what the agent is doing, why, how confident it is, and where the human can step in. Not interaction-by-interaction design. Supervision design.
Agent-to-human handoff becomes a core design problem. When the agent reaches its limits, the transition to a human has to preserve full context. The human receiving the handoff can't be expected to re-investigate what the agent already processed. This is handoff design at a different scale from what we covered in the design handoff evolution blog.
Non-deterministic behaviour needs to be framed as adaptation.Same input might produce different outputs depending on the agent's context, confidence, and learning history. Users will encounter inconsistency. The UX has to communicate why the agent made a different decision this time instead of leaving the user confused.
Failure messages need substance. "Something went wrong" doesn't work for a system that was operating on the user's behalf without supervision. The failure message has to say what the agent was trying to do, where it stopped, what it completed before the failure, and what the user needs to do next. Failure design is one of the core UX skills AI can't replace.
Autonomy level selection has to be a first-class control.Not a settings page buried three menus deep. A visible, adjustable control that moves from "check with me before everything" to "handle it, show me a summary." Smashing Magazine calls this "the autonomy dial" [3]. The user adjusts it per task, per context, per risk level.
The Automation Boundary Is a Design Decision
Once you can automate more things than before, the temptation is to automate all of them.
Can you set up an agent to calculate payroll by reading timesheet data and transferring money automatically? Technically, probably. Should you? Not without a human reviewing every pay run. Edge cases, exceptions, adjustments, legal requirements, and context-dependent decisions make full autonomy unreliable for something this consequential.
The designer's job on an agentic project is to draw the line between what benefits from autonomy and what requires human judgment. Which tasks can the agent handle alone? Which ones need a human check? Where does the agent stop and ask? Where does it proceed? These are UX decisions, and they carry more weight than any interaction pattern you'll design on the project.
Getting the automation boundary wrong in a traditional interface means a button misbehaves. Getting it wrong in an agentic system means the agent does something irreversible with the user's money, data, or relationships.
We're building the capability to design for these systems at Xperience Wave. The core design skills we've always focused on, judgment, stakeholder influence, ethical reasoning, systems thinking, apply here more than anywhere else in design right now. Book a strategy call if you want to talk about where agentic design fits in your career.
Free Resource: The 10 Principles of Agentic UX
A one-page reference covering all 10 principles with the specific design pattern for each: transparency of reasoning, confidence signaling, the autonomy dial, intent preview, interruptibility, action audit trails, handoff design, multi-agent visibility, adaptation visibility, and graceful degradation. Plus the three human involvement levels (in/on/out-of-loop) with when to use each.
Print it. Pin it to your desk. Pull it up before your next agentic design review.
The 10 Principles of Agentic UX
All 10 principles with the specific design pattern for each, plus the three human involvement levels and when to use each. One page. Printable.
No spam. Just the principles.
Sources & References
- [1] Bandi et al. / arxiv. (2026). "Human-AI Agent Interaction in a Business Context." Mixed-methods study developing UX principles for human-AI agent interactions. arxiv.org
- [2] UX Raspberry / Medium. (2026). "The Agentic Interface: Principles and Patterns for Autonomous User Experiences." medium.com
- [3] Smashing Magazine. (2026). "Designing For Agentic AI: Practical UX Patterns For Control, Consent, And Accountability." smashingmagazine.com
- [4] Mitchell, M. et al. (2025). "Fully Autonomous AI Agents Should Not Be Developed." arxiv. arxiv.org
- [5] Capgemini Research Institute. (2025). "Rise of Agentic AI: How Trust Is the Key to Human-AI Collaboration." Trust in fully autonomous agents dropped from 43% to 27% in 12 months. capgemini.com
- [6] International AI Safety Report. (2026). "Automation Bias Persists With New AI Tools." arxiv.org
- [7] Designative. (2026). "Trust Calibration in Agentic AI: Designing for Appropriate Reliance, Not Blind Trust." Referencing Buçinca et al. (CHI 2021), Parasuraman & Riley (1997), Bainbridge (1983). designative.info
- [8] Prabhakar, R. (2026). "10 Real-World Agentic AI Examples That Are Actually Generating Revenue in 2026." rohitprabhakar.com
- [9] LiqTeq. (2026). "15 Real-World AI Agent Use Cases Across Industries." liqteq.com
- [10] AIMultiple. (2026). "40+ Agentic AI Use Cases with Real-life Examples." aimultiple.com
- [11] AlphaBold / Gartner. (2026). "Top Agentic AI Use Cases That Actually Pay Back in 2026." Gartner expects 40% of enterprise apps to include agents by end of 2026. alphabold.com
- [12] AufaitUX. (2026). "Agentic AI Design Patterns: Enterprise Guide." 10 recurring patterns including Planning, Reflection, Tool Integration, Multi-Agent Collaboration, Trust Calibration. aufaitux.com
Further Reading on Xperience Wave
- AI Will Change Everything About Design. Except the Part That Actually Matters.
- 5 Things Senior Designers Should Be Doing With AI - None of Them Involve Figma Plugins
- From Slicing PSDs to Shipping Front-End Code: How Design Handoffs Evolved - And What Must Never Change
- How to Build Your Personal AI Workflow as a Designer (Without Losing What Makes You Good)
About the Author
Shaik Murad is Co-founder and Head of Product & Design at Xperience Wave, a UX design studio based in Bangalore, working across designer mentorship, UX services for businesses, and a design community of 1,370+ designers.
- Murad, Co-founder & Head of Product & Design, Xperience Wave
