The “everything becomes a chat box” prediction sounds elegant until the agent wants to change a payment, publish a page or email the wrong person.

The appeal is obvious: an agent could complete a task without making someone click through five screens.

That argument sounds plausible until the agent is doing something expensive, irreversible, sensitive or ambiguous.

Then the interface suddenly matters again.

Agentic systems shift the user experience from direct manipulation toward delegation. Instead of completing every step, the user expresses an outcome and the software works through the process.

But delegation creates new design requirements.

What exactly is the agent about to do? Which systems can it access? What information will it send? Which actions require approval? What happens if it is wrong? Can the user inspect or reverse the result?

These are interface questions.

The future UI may therefore have fewer repetitive controls but more surfaces for intent, visibility and control.

A useful agentic interface might show:

• the goal it believes it is pursuing
• the steps it plans to take
• the data and tools it will use
• checkpoints requiring approval
• progress while work is underway
• exceptions or uncertainty
• the final artefact
• a reversible history of actions

This is closer to supervising a capable operator than filling in a form.

Figma’s MCP tools now let agents read design context and, in supported workflows, write native canvas content. The canvas still gives people a place to inspect and refine the result.

The same principle applies to business applications.

ERP systems, CMS platforms and operational tools are not likely to become a single conversational window. Their state matters too much.

A sales manager needs to see pipeline structure. A content editor needs publication status. A finance team needs ledgers and approvals. AI may perform more of the underlying work, but humans still need reliable representations of the system.

So the interface changes role.

Yesterday’s UI was primarily an instruction surface.

An agentic UI can become a supervision surface as well as an instruction surface.

That is a significant shift for designers.

The question is no longer just, “How does the user complete this workflow?”

It becomes, “How does the user understand, direct and trust a workflow that software may complete on their behalf?”

That is not the death of UX.

It is an expansion of UX into agency, accountability and system behaviour.