Capability discovery in agentic UX

In a traditional interface, buttons advertised themselves. The user didn’t have to guess that price filtering was possible; “Filter by price” was on the screen. The interface was its own menu of capabilities.

In an agentic experience, this disappears. The user looks at a text box and freezes. What can I ask for? How much can I ask for? In what tone? The AI has incredible capabilities, but if the person doesn’t know they exist, they don’t use them. And if they don’t use them, the product fails.

This is one of the biggest UX challenges for AI products today. It’s called capability discovery, and there are concrete patterns that help.

Why the problem is bigger than it looks

People who work in AI live in a bubble. They know GPT can generate 3D files for printing, that Claude can code an entire app from a conversation, that they can ask for a complete travel plan from Lisbon to Tokyo.

Most of your users know none of this. Parents, colleagues outside the bubble, some friends: they arrive at a chatbot and use it like Google. Short question. Short answer. Frustration when the answer isn’t the one they expected.

The problem worsens in multi-agent. If the system has 5 specialist agents, each with 10 tools, that’s 50 latent capabilities no one will discover by themselves. Expecting the user to “explore” is, in practice, guaranteeing that 90% of capabilities never get used.

Five patterns that work

1. Tasks as examples. Instead of an empty text box, give 3 to 5 typical starter tasks. “Plan a trip to Lisbon”, “Summarise this PDF”, “Compare these two files”. The user clicks one, sees the result, gets a feel for the kinds of things to ask. It’s like a menu, but curated.

2. Proactive contextual suggestions. During the flow, suggest the next step. When someone finishes sharing a document with the AI, surface “I can summarise, translate, or extract the main points”. It’s not spam: it’s capability discovery embedded at the moment it makes sense.

3. Light documentation at the right time. Not a separate help centre. A line that appears when relevant: “I can also do this in other languages, just ask”. Nielsen’s “help and documentation” heuristic still holds. It applies differently: contextual documentation instead of a manual.

4. User history as a hint. If the person used a rare capability before, remind them later. “Last time I also did X. Want me to?” Turns historical behaviour into implicit education.

5. Show what others have asked. Without breaking privacy, aggregate common requests in an examples section. “People often ask…”. Helps newcomers and signals the system is capable.

The trade-off: capability vs. simplicity

The designer’s instinct is to minimise the interface. Clean box, short prompt, let the AI talk. Works when the user already knows what to ask for. Fails when they don’t.

The other extreme, listing every capability in a vertical menu, drags us back to the traditional interface and loses the fluidity of conversation. Worst of both worlds.

The middle ground that’s been working: start with visible contextual suggestions, then taper as the user gains confidence. Power users end up preferring the empty box. Beginners need the rails. It’s not “either-or”, it’s progression.

What this asks of multi-agent

When there are multiple agents, capability discovery becomes a double problem. First, the user needs to know what the whole system can do. Second, ideally, understand which agent is handling what (attribution, covered in Multi-agent orchestration).

An elegant solution is to surface capabilities aggregated, not by agent. The user doesn’t need to know there’s a “search agent” and a “comparison agent”. They need to know they can search and compare. The internal split stays as a technical detail.

How to measure capability discovery

Three metrics I track:

1. Tool coverage per session. How many of the available tools were used? Low number means invisible tools.
2. Time-to-first-non-trivial-action. How long until the person asks for something non-trivial (a simple question)? High means capability discovery is failing.
3. Repeated request patterns. People ask 3 or 4 things and stop? Sign they’ve discovered a small subset and stopped exploring.

Where this fits in the rest

Capability discovery is one of the three core pieces of agentic UX, alongside Observability and cost communication. The general background is in the Design for AI guide.

Nielsen’s “help and documentation” heuristic remains a good lens. The difference is that in agentic products, help and documentation aren’t separate pages: they’re micro-moments during use. Worth revisiting the 10 heuristics with AI eyes when designing for this.