What Are AI Agents?

AI agents are autonomous software systems that can perceive their environment, make decisions, and take actions to achieve specific goals - often with minimal human intervention. Unlike traditional software that follows rigid, pre-programmed rules, agents can interpret context, adapt their approach, and work through multi-step tasks by reasoning about what needs to happen next.

Agentic AI refers to this new category of AI systems that exhibit agency: the ability to act independently, plan sequences of actions, use tools, and learn from feedback. Rather than simply responding to a single prompt, agentic systems can break down complex objectives, iterate on their approach, and coordinate across multiple interfaces and data sources.

In practice, this means:

• An agent might read an email, understand the request, pull relevant data from multiple systems, draft a response, and schedule a follow-up - all without step-by-step human guidance.
• Agents can operate across traditional UIs (clicking buttons, filling forms), APIs, databases, and documents, treating the entire software landscape as their working environment.
• They're designed to handle ambiguity, recover from errors, and ask for help when needed, making them more resilient than traditional automation.

The shift to agentic AI is significant because it changes who - or what - interacts with our software. Interfaces originally built for human eyes and hands now need to serve both people and autonomous agents working alongside them.

Over the past year, the contract between people and software has begun to shift. Interfaces are increasingly read and acted upon by software, not just clicked by humans. That suggests making intent, structure, and constraints more legible, so both people and agents can find their way.

Recently, I noticed how AI agents interact with software: some struggle with basic tasks whilst easily completing more complex ones. What if we designed interfaces that work well for both humans and agents?

‍

Designing UI That Agents Can Notice

This is increasingly being referred to as UX/AX - User Experience and Agent Experience. The experience is no longer limited to humans. Agentic systems also need to be considered if we want AI‑friendly applications that bring efficiency to modern workflows.

Agents infer a lot from names, layout, and metadata. When a UI is ambiguous to humans, it is often opaque to agents.

• A small example: a calendar that reveals an add button only on hover feels tidy, but it is hard for an agent to discover. An always‑visible "Add event" button is straightforward for both.
• Clear naming helps: "Customer status" is clearer than "CustStat," and labels like "Units" need context.
• Structure and field grouping help too: separating supplier name and address from receiver name and address makes state and relationships clearer.
• Light annotation carries weight: hints and placeholders can guide formatting, especially for dates with different formats such as DD/MM/YY.

‍

Keeping People in the Loop

Predictability and reversibility build trust.

• Preview what will happen, what data will be touched, and the expected outcome to reduce anxiety.
• Provide reversible actions, version history, and clear reconciliation steps so it is safe to experiment.
• When confidence is low, adapt the UI: show missing‑data flags or uncertainty and offer safer choices.
• Add a simple "Why this?" to reveal inputs, rules, and references.

‍

What Products May Need to Offer Agents

Think of the product like an integration guide for a teammate. In logistics, you often need domain knowledge to parse acronyms and flows.

• Clear verbs: "Create record," "Submit," "Search {X}."
• Observable state: Draft, Pending, Submitted, Accepted, Held, Withdrawn, with timestamps and who set them.
• Predictable actions: preconditions, postconditions, and expected side effects.
• Prioritise clarity in the flow over fewer clicks.

‍

Treating Knowledge as Part of the System

Agents are only as good as the information they can access and trust.

• Maintain a shared glossary of domain terms to keep everyone aligned.
• Document standard procedures with clear preconditions and validation steps.
• Define compliance requirements, exception handling, and escalation paths up front.
• Provide concrete examples of both correct and incorrect approaches.
• Organise documentation into focused, titled sections with stable references. Use lists and tables to make relationships clear.
• Surface relevant context in the UI itself - current page, selected records, active filters - so agents can cite their sources.

‍

A Few Anti‑Patterns

• Chat‑only interfaces for operational work
• Hidden state in free‑text notes
• Overloaded fields with multiple meanings
• Magic buttons with unclear side effects
• Labels without units or validation
• Multiple entry points for the same critical action

‍

What to Watch

It may be helpful to track outcomes rather than just activity.

• Task completion rate without human intervention
• Plan approval versus rejection rate
• Time to first correct action
• Rework and rollback rate

‍

Where This Might Be Heading

Agentic AI does not eliminate UI - at least not yet. If anything, it nudges us towards clearer, more explicit interfaces. Natural language can be great for capturing intent, whilst structured UI supports precision and repeatability. Designing for two users - a person and an agent - may produce systems that feel more intuitive, resilient, and fast.

Curious to hear from designers and developers working in this space: What patterns are working for you? What is breaking? And do you think UX/AX is here to stay?

‍