UI... Where We Are Going We Don't Need UI

Recently, on a podcast when discussing building I said, “Please get your UI out of my face. I want my agent to do everything.” And then I explained that, yes, at work we are still building UIs. Which is a very normal thing to find yourself saying as a product person these days. It's messy out here in the world of AI and product.

In my own work, I've moved into the terminal, where I can give an agent instructions and work with it directly. That's where I feel I have the most control over what the AI is doing. I want to tell it to go figure something out and report back. I don't want to open another application and click through a bunch of screens.

I also build products for people whose work looks different from mine. We also sell to enterprise and those are slow moving ships. While I'm out here living in the future, I have a sales team presenting to decision makers who still want the clicky-clicky stuff like it's 2016.

One product I currently work on is an AI router/gateway. This product gives an organization a way to manage access to AI models and see how they're being used. A developer wants to try a model, get it working in their application, and keep moving. Somewhere else in that organization, someone is responsible for understanding what happens when a hundred developers do that. Or a thousand.

Who's spending the money? Where is the data going? Does that match what they've promised their customers? Can they put some sort of limit on a LLM before an agent spends the afternoon trying very hard to solve a problem it cannot solve?

Those are all things the product has to help someone handle.

Think about an administrator opening a dashboard and seeing 10,000 API keys, the credentials applications use to access the models. They can look at which user a key belongs to, which region it's tied to, and how much is being spent. They can start with the whole organization and then look more closely at a particular user.

There's a reason for that screen to exist. Someone has a job to do with it. My enthusiasm for asking an agent to handle my own work doesn't tell me whether that administrator wants to do theirs the same way and if they did, are they even allowed to within that organization?

Like I said an enterprise sales still needs the clicky-click demo. People want to see what they're buying and understand how they would use it. My working hypothesis is that we need to keep building that experience while also figuring out how a developer's agent should interact with the product.

How do I balance wanting the UI out of my face with trying to predict where AI is going.

For me, this is all just product work. What problem are we trying to solve? Who can we talk to? What does the data tell us? And, please, is the data clean enough to tell us anything useful?

I spend a lot of time reading, having conversations, and trying these tools myself. AI helps me work through my notes and look for themes across what I'm hearing. Those themes give me things to investigate. Then I can form a hypothesis and make a small bet, with some idea of what I need to learn from it. Like I said, just product work.

So yes, I'm thinking about how someone will point their agent at our product and ask it to get something done. I'm also thinking about what an administrator needs to see when they open the dashboard. Both experiences need product work, and right now even those loads of us would like to transcend the UI, we still have reasons to build both.