I'm coming to believe that what separates successful AI agents isn't intelligence. Foundation models are smarter than we need. It's UI that helps users review and apply the agent's output.
Consider the most successful AI integrations so far: customer support agents and coding agents. Good UI patterns existed for both and are ready for wide adoption.
Customer support agents don’t need much UI invention. It’s a familiar chat interface. The occasional buttons in the chat existed before the recent AI emergence.
Coding agents are similar. For example, take Claude Code. It’s a chat UI with git diffs. Engineers work with patches all the time. Code suggestion in IDE existed for a while. It felt novel and natural to work with.
Another advantage of coding agents is that engineers work in plain text editors. It’s easier to show diffs in document files. Established patterns exist to show suggested changes like commenting and spell checking.
We don’t have a best practice to show AI agent changes in the common application UI. When automating a task, there are typically 3 steps.
- Your data — whether imported from third parties or in your application for the customer.
- Agent work — the agent takes your data and makes some changes.
- Review the change — go back to 2 if necessary or finish if all looks good.
We lack a good UI pattern in 3, and that’s impacting many applications. Successful cases like customer support and coding agents have this. All “AI” companies end up becoming a chat app in one way or another because that’s the only way you skip the invention of 3. Its AI gives you the answer and you take action — leaving manual work to humans.
This UI gap shapes who benefits from AI today. Without good review interfaces, humans still do the final step — and that step requires domain expertise. Knowledge workers adopt AI first not because their jobs are easier to automate, but because they can handle the manual review that bad UI forces on them.
We need UI inventions to show AI agents’ changes in natural ways in common application interfaces like forms and tables. That will unlock huge value creation in the decade of agents.