There is a wall. It is vast, concrete, unbroken. It stretches in every direction - the wall of autonomous operation, of agents acting in the world without human involvement. And in this wall, there is a slit. Narrow. Precisely cut. Through it, a sliver of golden light falls into the space where the human stands.
This slit is the Agent-to-User Interface. It is not a dashboard. It is not a control panel. It is not a conversation. It is the narrowest possible opening through which an autonomous system communicates with the human who delegated authority to it - and through which the human can, when necessary, reach back in.
In screen-based design, the interface was everything. The entire experience was the interface. Every pixel, every interaction, every moment of attention was mediated through a designed surface. The designer's job was to make that surface as rich, as intuitive, as engaging as possible. More interface meant more opportunity for design.
In agentic design, the opposite is true. The best agent-to-user interface is the one that appears least often. The most successful agentic experience is the one where the slit in the wall is never needed - where the agent operates with such fidelity to the human's intent that no interruption is required. The interface is not the experience. The interface is the exception.
What is A2UI
A2UI - Agent-to-User Interface - is a term I propose for the specific, designed surface through which an autonomous agent communicates with the human principal who delegated authority to it. It is distinct from a traditional user interface in three fundamental ways.
First, it is initiated by the agent, not the human. In traditional UI, the human opens the application, navigates to a screen, and interacts with elements the designer has placed there. In A2UI, the agent determines when communication is necessary and surfaces the interface at that moment. The human does not go to the interface - the interface comes to the human.
Second, it is transient by design. A traditional interface persists - it is always available, always present, always waiting for the human to return. An A2UI appears when needed and disappears when resolved. It has no resting state. Its natural condition is absence.
Third, it is high-stakes by definition. If an agent surfaces an A2UI, it means the agent has encountered a situation that exceeds its delegation scope, requires human judgment, or involves a consequence significant enough to warrant interruption. Every A2UI appearance is a signal that something important is happening. There are no casual A2UI interactions.
"The A2UI is not a dashboard you check. It is a fire alarm that sounds only when there is an actual fire - and the design challenge is ensuring that when it sounds, you trust it, understand it, and can act on it immediately."
The Interrupt Surface
Every A2UI is, at its core, an interrupt. It breaks the human's absence - their state of not-attending-to-the-agent - and demands attention. This makes interrupt design the central discipline of A2UI, and it is a discipline that UX has never had to master at this level of consequence.
In screen-based design, interrupts were notifications - and notifications became a design disaster. The notification economy trained users to dismiss, ignore, and silence. Push notifications, badge counts, toast messages, modal dialogs - each began as a useful signal and degraded into noise. By 2024, the average smartphone user received over 80 notifications per day, and the tap-through rate on most categories had fallen below 2%.
A2UI cannot afford this degradation. When an autonomous agent managing your pension portfolio surfaces an interrupt to say that market conditions have triggered a rebalancing threshold that exceeds your pre-authorised parameters, you need to attend to that interrupt. If the A2UI has been crying wolf - surfacing trivial confirmations, unnecessary status updates, low-stakes acknowledgments - the human will have been trained to dismiss it. And the one interrupt that matters will be lost in the noise.
Five Modalities of A2UI
Not all agent-to-user communications are equal. I propose five distinct modalities of A2UI, each with different design requirements, urgency levels, and trust implications.
I. The Alert
The highest-urgency modality. The agent has encountered a situation that requires immediate human decision. The delegation scope has been exceeded, a threshold has been breached, or a consequential error is imminent. The Alert demands attention now. It must be unmissable, unambiguous, and actionable within seconds. In banking: "Your agent has been offered a fixed-rate mortgage at 4.2% that expires in 90 minutes. This exceeds your pre-authorised rate ceiling of 4.0%. Approve, reject, or modify ceiling."
II. The Checkpoint
A scheduled or milestone-triggered pause in autonomous operation. The agent has reached a pre-defined decision point and is requesting confirmation before proceeding. Unlike the Alert, the Checkpoint is expected - the human knew it was coming. It is a designed moment of human re-engagement in an otherwise autonomous process. In insurance: "Your agent has completed the annual review of your home insurance. Three alternative policies were evaluated. The recommended switch saves £340/year with equivalent coverage. Review the comparison before your agent proceeds."
III. The Digest
A periodic summary of autonomous activity. The Digest does not require immediate action - it is an observability mechanism that allows the human to maintain awareness of what the agent has been doing without monitoring it in real time. The design challenge is making the Digest informative without being overwhelming, and honest without being alarming. In wealth management: "Weekly digest: Your agent executed 14 transactions this week. Portfolio allocation shifted 2.3% toward bonds following your risk-reduction preference. Net performance: +0.4%. No threshold breaches. No anomalies."
IV. The Handback
The agent is returning authority to the human. This is not a failure - it is a designed boundary. The agent has reached the edge of its competence, its delegation scope, or its confidence threshold, and is gracefully transferring control back to the human with full context. The Handback is the most trust-critical A2UI modality, because it is the moment where the agent demonstrates self-awareness of its own limitations. In legal: "Your compliance agent has identified a regulatory change that may affect your data processing agreements. The change involves jurisdictional nuances that exceed the agent's confidence threshold. Full analysis attached. Human review recommended before the agent updates any compliance documentation."
V. The Ambient Signal
The lowest-urgency modality - and arguably the most innovative. The Ambient Signal communicates agent status without demanding attention. It is the equivalent of a pilot light: visible if you look, invisible if you do not. It answers the question "Is everything okay?" without requiring the human to ask. In smart home: a subtle colour shift in a light strip that indicates your energy agent is in active negotiation with the grid. No action required. No attention demanded. But the information is there if you want it.
"The five modalities form a spectrum from maximum interruption to zero interruption. The designer's task is to ensure that every agent communication is assigned to the correct modality - because an Alert delivered as a Digest is a catastrophe, and a Digest delivered as an Alert is an erosion of trust."
The Notification Problem
The history of digital notifications is a cautionary tale for A2UI designers. Every notification system in the history of computing has followed the same arc: useful signal becomes tolerable noise becomes intolerable spam becomes silenced channel. Email notifications. SMS alerts. Push notifications. Badge counts. Toast messages. Every single one.
The reason is structural, not incidental. Notification systems are designed by the sender, not the receiver. The incentive structure rewards more notifications, not fewer. Every product team believes their notification is the important one. The result is a tragedy of the commons: each individual notification may be justified, but the aggregate effect is a communication channel that no one trusts.
A2UI must break this pattern. And the way to break it is to invert the incentive structure. In A2UI, the agent is penalised for unnecessary interruption, not rewarded for engagement. The metric is not "did the human see the notification" but "did the human need to see the notification." Every interrupt that was unnecessary is a withdrawal from the trust account. Every interrupt that was necessary and well-timed is a deposit.
This requires what I call interrupt budgeting - a design practice where the agent has a finite budget of interrupts per time period, and must allocate that budget to the communications that matter most. If the budget is five interrupts per week, the agent must triage ruthlessly. This constraint forces the agent - and the designer - to distinguish between what is truly important and what merely feels important.
Contextual Surfacing
When an A2UI does appear, it must appear in the right context. This means not just the right time but the right place, the right device, the right cognitive state, and the right emotional register.
Consider a financial agent that needs to surface a Checkpoint about a significant investment decision. If the human is driving, the A2UI should not appear on their phone. If the human is in a meeting, the A2UI should queue until they are available. If the human is about to go to sleep, the A2UI should assess whether the decision can wait until morning - and if it can, it should wait.
This is contextual surfacing: the practice of delivering agent communications not just when they are relevant but when the human is in a state to receive them well. It requires the agent to have a model of the human's availability, attention, and cognitive load - not to surveil them, but to serve them. The distinction between surveillance and service is one of the most important ethical boundaries in A2UI design.
The design pattern I propose is urgency-weighted queuing. Every A2UI communication is assigned an urgency score and a decay rate. High-urgency communications (Alerts) have a short decay - they must be delivered soon or they lose relevance. Low-urgency communications (Digests) have a long decay - they can wait for the optimal moment. The queue is processed when the human's context is receptive, with higher-urgency items surfaced first.
The Absence Principle
The most radical aspect of A2UI design is what I call the Absence Principle: the recognition that the default state of the agent-to-user interface is non-existence. The interface does not exist until it needs to exist. It has no home screen, no landing page, no persistent navigation. It is pure potential - a surface that materialises when required and dissolves when resolved.
This is profoundly disorienting for designers trained in screen-based thinking. In UX, the interface is the product. In A2UI, the absence of the interface is the product. The agent's success is measured by how rarely the human needs to see the interface, not how often they engage with it.
But absence is not emptiness. The human must feel confident that the agent is operating correctly even when no interface is visible. This is the observability paradox at the heart of A2UI: the human needs to know that things are fine without being shown that things are fine. The Ambient Signal modality addresses this partially, but the deeper solution is trust architecture - the accumulated evidence, over time, that the agent acts with integrity when unsupervised.
"In screen-based design, the interface is the product. In A2UI, the absence of the interface is the product. The agent's success is measured by how rarely the human needs to see the interface."
Designing the Return
There is a complementary surface to the A2UI that deserves its own design attention: the return interface. This is the surface the human encounters when they voluntarily choose to check on the agent - not because the agent interrupted them, but because they want to know what has been happening.
The return interface is fundamentally different from the A2UI. It is human-initiated, not agent-initiated. It is exploratory, not urgent. It is a moment of curiosity or mild anxiety, not a moment of decision. And it must be designed to reassure without patronising, to inform without overwhelming, and to offer depth without requiring it.
I propose a design pattern called progressive disclosure of autonomy. When the human returns, the first layer shows the outcome: "Everything is fine. Your agent completed 23 actions this week. Net result: positive." The second layer, available on request, shows the summary: what categories of actions were taken, what decisions were made, what the agent's confidence level was. The third layer shows the full audit trail: every action, every decision point, every piece of reasoning.
Most humans will never reach the third layer. But its existence is essential. It is the architectural foundation of trust - the knowledge that you could see everything, even though you choose not to. The return interface is the agent saying: "I have nothing to hide. Look as deeply as you wish."
A2UI in Banking
Banking is the sector where A2UI design will be tested most rigorously, because banking agents handle the most consequential delegations: money, contracts, financial commitments that extend years into the future.
Consider a mortgage agent. The human delegates authority to find, evaluate, and negotiate a mortgage. The agent operates autonomously for weeks - scanning the market, comparing rates, assessing terms, negotiating with lenders. During this period, the A2UI might surface a single Digest per week: "12 products evaluated. 3 shortlisted. Negotiation in progress with 2 lenders." Then, at the critical moment, a Checkpoint: "Your agent recommends Product A at 3.89% fixed for 5 years. Product B offers 3.82% but with an early repayment charge that conflicts with your stated flexibility preference. Review and approve."
The entire mortgage journey - weeks of autonomous operation - is mediated through perhaps five A2UI moments. Five slits in the wall. Five narrow openings through which the human exercises oversight of a process that, in the pre-agentic era, would have required dozens of hours of personal attention.
The design quality of those five moments determines whether the human trusts the agent, whether they feel in control despite being absent, and whether they will delegate again. Five interactions. Five opportunities to build or destroy a relationship. This is why A2UI design is the highest-stakes interaction design in the field.
The Narrow Opening
The image I began with - the slit in the concrete wall - is not merely metaphorical. It captures something essential about the nature of agent-to-user communication in the agentic age.
The wall is necessary. It represents the boundary between human attention and autonomous operation - a boundary that must exist for agentic systems to deliver their value. If the wall is transparent, if the human can see everything the agent does in real time, the agent becomes a tool, not an agent. The human is back in the loop, monitoring, supervising, second-guessing. The value of delegation evaporates.
But the wall cannot be solid. A wall with no opening is a system with no oversight - and a system with no oversight is a system that cannot be trusted. The human must be able to reach through when they need to, and the agent must be able to reach through when it needs to.
The slit is the design solution. Narrow enough that it does not compromise the wall's integrity. Wide enough that meaningful communication can pass through. Precisely positioned so that the light falls where it is needed. This is A2UI: the narrowest possible opening between human attention and autonomous action, designed with the precision and restraint that the highest-stakes interactions demand.
The designers who master A2UI will not be the ones who build the most beautiful dashboards or the most engaging notification systems. They will be the ones who understand that in the age of autonomous agents, the greatest design achievement is the interface that almost never appears - and is absolutely trusted when it does.
"The greatest design achievement in the agentic age is the interface that almost never appears - and is absolutely trusted when it does. That is A2UI. That is the narrowest possible opening."