Agentic UX Predictions

Hi all,

Lately there’s been one topic that has dominated our conversations: How will user experiences change with the rise of AI Agents?

Wanting to address this topic in more depth I’m breaking with the usual format of a few vignettes for this issue and instead offer one long-ish form piece. I hope you get lots of actionable insights and enjoyment from it.

We’ve crossed a Rubicon. With the advent of AI Agents, we are entering an era in which AI is making complex decisions and actions that were previously only made by humans.

This shift signifies a move towards autonomous systems capable of not just assisting but replacing specific human roles. This evolution necessitates a complete rethink of product design, user roles, and control mechanisms.

Much like the transition from command lines to visual interfaces, the move to agentic AI presents an opportunity to reshape how we interact with technology. Understanding this shift is crucial for navigating the future of AI-driven products.

Traditional AI systems are designed to augment human capabilities, providing data, insights, and support to help us do our jobs better.

Agentic AI marks a departure from this model. It’s about creating systems that can execute tasks and make decisions autonomously for a given vertical. As Jared Friedman, Managing Partner at Y Combinator, put it, it’s like a traditional SaaS product and its users combined into a single entity.

In general, this is best illustrated by OpenAI’s Operator, a chatbot that can perform tasks using your browser just like you would. You ask it what you need it to do, and it will break down the request, infer the best course of action, and then carry it out right before your eyes.

Operator is just the beginning. Soon, we’ll see entire enterprise functions effectively run by Agents. This begs a few questions: What does an agentic AI product need to be capable of? Who will be its users when AI replaces entire departments? And how do you design for this world to create a market-leading solution?

For product companies at the forefront of this development, an unusual question presents itself: If your product replaces those who would have previously been its users, who are you designing for now?

Initially, we’ll see an increased demand for subject matter experts (SMEs). As AI products still produce pitifully inconsistent and underwhelming results, these users will be responsible for configuring, troubleshooting, and refining AI systems and their outputs.

These SMEs may not necessarily be AI engineers but expert marketers, analysts, designers, and recruiters, all at the top of their game and able to tell when a product isn’t working quite right. These users will need to be able to tweak configurations for optimal performance, encoding their implicit knowledge into a tool for better and more consistent outcomes.

Today, the prime example of this is software engineers. We are already seeing relatively mundane development projects fully automated by AI-powered products. However, these are far from perfect, and they still run into many errors or do not deliver solutions that are at their maximum efficiency.

Those who require reliable delivery of complex code projects can not dispense with a human backup, someone they can turn to to resolve roadblocks the AI tool alone can not overcome. This principle will continue to apply to any expertise at a certain level of complexity and depth.

However, as agentic AI systems mature and become more sophisticated, the specialised expertise required to manage, maintain, and troubleshoot them will gradually decrease. This leads to the second major shift in user roles, where we will see everyone turn into a manager.

AI Managers will primarily focus on defining strategic outcomes and setting parameters for the AI systems. They will make key decisions, such as balancing budget, speed, and quality.

How pronounced and rapidly these roles will manifest in each industry can already be gleaned by considering AI adoption statistics today.

Competition among tech companies in each vertical will be fiercer, and the need for their products to interface with one another will be more pronounced, as user expectations will be for software to know and do everything.

There will be dedicated agents for every aspect of business, but the user won’t necessarily want to care. Context switching kills adoption, which means we’ll likely see a few AIs take the ‘operating system’ role and provide users access to all capabilities.

Those who establish themselves as the go-to ‘app’ for integrators and empower users to be effective AI Managers will win.

The evolution of agentic AI is not happening in a vacuum. We can draw valuable lessons from the democratisation of other technologies before it.

Web development has transitioned from complex coding to user-friendly platforms. Initially, creating a website required extensive knowledge of design and development. Then came Squarespace and Wix, enabling users to build basic, generic, and informational websites with minimal effort. However, as businesses needed more control and flexibility, platforms like Webflow emerged. While Webflow has a steeper learning curve, it offers greater creative freedom, robust content management systems, and seamless integrations with other business applications. Now, Webflow itself is on a bath of being outdone by Framer, which does away with confusing ‘developer lingo’ and offers similar capabilities with increased accessibility and a flatter learning curve.

3D modelling was once reserved for experts. Then, tools like SketchUp and Spline made it far more accessible. Beyond ease of use, these platforms introduced integrations that allowed 3D models to be used across websites, mobile apps, augmented reality, and rapid prototyping. The value was not just in accessibility but also in giving users more control over where their creations were used and how well they could integrate their 3D modelling tool with other systems.

Video editing has also evolved significantly. Traditional tools like Adobe Premiere were powerful but time-consuming. CapCut made editing more accessible, but true innovation came with Descript, which introduced AI-powered, text-based video editing. This shift did not merely simplify video editing—it redefined users’ control over their content through automation.

These examples illustrate a universal trend towards more user-friendly versions of previously complex and inaccessible tools, which can be seen in every area of business, from the generation of insights to communication, IT, finance, HR, customer research, and more.

Crucially, the removal of complexity was achieved not by removing control but by making it easier for users with less specialised training to control the outcomes.

While ‘quick and easy’ alternatives exist—like libraries of website templates, pre-made 3D models, or consumer-friendly video editing tools—users often reject them if they compromise control and quality.

Indeed, a lack of proper controls has hampered the adoption of generative AI tools to date. In turn, the most meaningful advancements of these products have been in this area:

Using sample images, say, to ‘fix’ an image’s style or geometry (Adobe Firefly), introducing key frames to make video generation more predictable (Krea.ai), improving character consistency for visual media, or sampling sounds for audio generation.

In business, advancements in RAG (Retrieval Augmented Generation) methods have produced outputs that are increasingly relevant to a particular organisation or domain, as has the maturing of guardrails to constrain these outputs.

The need for control balanced with accessibility is inherent to all viable future solutions. As long as technology delivers something we care about, we’ll want to exert a degree of control over its outputs. The question is not how to automate this but to make it as seamless and accessible as possible.

As AI capabilities continue to evolve, the user’s role will shift towards greater strategic oversight in combination with accessible ways of controlling outcomes. In other words, Micromanagement. To see what this looks like in practice, let’s consider the core needs of a micromanager:

Priority-setting frameworks in management abound, and many have visual representations of their workings. Once you’ve found the one that works best in your context, these frameworks can directly inspire the appropriate user interface.

For instance, let’s take the speed/quality/cost triangle, which finds its way into discussions with our partners almost daily. Products will be designed around these intuitive control frameworks, which allow influencing AI Agents without the need for a deep understanding of their underlying technical complexities.

Controls on input measures are only helpful and instil trust and confidence if they are combined with accurate feedback on their impact. Agents must proactively assist users in predicting the consequences—intended or unintended—of their proposed actions to help them anticipate and protect critical outcomes.

This Scenario Planning step is where most of the human-in-the-loop interactions will occur. Proactive suggestions for optimising the system will be made, but the user needs to have the final say on locking in minimum thresholds or protecting certain outcomes.

AI tools today are prone to overwrite previously agreed outcomes while responding to new situations. Agentic systems will only be ready for use in business once this quirk - alongside frequent hallucination - has been eradicated and users can ‘lock’ essential levers.

In business-critical functions, users need to retain the ability to understand how different operations influence one another and how sets of agents interact. In many domains, users will learn to trust Agentic AI systems by virtue of their consistency in providing the right kinds of outputs. However, in enterprise applications, the ability to explain and tweak how an outcome is achieved remains paramount.

OpenAI’s Operator works on text-based chain-of-thought reasoning. As it performs its magic, you can follow this chain in the form of text that’s typed out on your screen as the AI makes deductions that inform actions and so on.

For scalable enterprise scenarios, this pattern has to level up in a few ways:

• What is done as a ‘one-off’ in Operator, with a user typing in prompts and observing results, needs to be scaled up so that the same workflow can be reproduced repeatedly and automatically with consistent results.

• To correct these links between a specific deduction and a resulting suggested action in these production systems we will need dedicated controls for every aspect of the whole system (more on this in the next section).

• This explainability pattern for text-based chain-of-thought will need to translate to other forms of information, most notably data analysis.

• Business strategies aren’t guided by only one goal, and they don’t only do one thing. In these complex systems, you will find not a linear chain but a network - a sort of ‘reasoning net’, where one deduction informs multiple actions and vice versa (Think, for example, how changing the production volume of a chocolate bar will impact absolutely everything in that business, from sales and the supply chain to revenue predictions and hiring and much more). Interrogating these relationships is not unlike a child continually asking “Why” each time they are given an answer, except there will be dozens and hundreds of answers each time.

How could users possibly navigate a system with such vast complexity?

As with every decent micromanager, a user will need to intervene from time to time. Much like you’ll sometimes want to change the font size in your beautiful website template from our historical examples above, there will always be details we’ll want to touch ourselves. However, the system described above is far too complex to recreate in a traditional interface with standard interfaces and fixed menus. We will need ephemerally augmented interfaces (EIs). Bear with me.

Important, frequently revisited, and more involved steps in a reasoning network may be locked and always available. Others may be queried and generated on the spot. Agents may anticipate and respond to a user’s intent and conjure relevant audiovisuals and controls—purpose-built for this moment.

For content creation agents in a video streaming service of the future, say, we may seek to ‘double click’ into assumptions about target demographics or tweak the style of visual effects for the next 200 episodes of the latest hit drama series.

In business, we may see impromptu data visualisations that give us tangible controls to move sliders and set limits.

Many applications and plugins do this to some extent, providing us with code interfaces, diagrams, visuals, differently formatted text, and editable canvases when they feel the context demands it. This type of impromptu interface will evolve to become the primary way of interacting with agents.

If some of the ideas discussed above seem otherworldly, consider how we might have felt about the state of AI five years ago. These shifts are imminent and lead to core tenets which product innovators should pay attention to when building Agentic AI software today:

• Accepting the dramatic reshuffle of the workforce and rethinking target users and use cases in this context

• Embracing simplicity (accessibility), transparency, and control as universal user needs

• Differentiating via a deep understanding of a given vertical and the people in it

• Learning from developments in the recent past to understand how to meet these needs in a specific domain

• Continuing to invest in deep user and customer insight to craft distinguished product value propositions and designing systems and interfaces that absorb complexity and transmit clarity

The emergence of complex, layered agent ecosystems will require sophisticated management strategies, which will form the frameworks for controlling them. Much like previous interface revolutions, the evolution of agentic AI will continue to be focused on making it universally accessible and malleable. It’s about empowering users to direct the outcomes while reducing the cognitive load required to manage AI-driven processes.

While technology may advance ad infinitum, it will continue to be created for humans. This means there will always be a ‘user’, even if who that is changes. And while users, technologies, and problems change, the product fundamentals of ‘who is the user and what problem do we solve for them’ will remain.