Give me a good API.

Most web applications assume that the web UI is the main place where work happens.

You log in.
You click around.
You find the right form.
You enter data.
You submit it.

Then you repeat that across five other systems just to go through one workflow. Writing software for example can involve the issue, some info from a wiki, the code itself, Claude, and then you just want to see if your colleague is available for a quick rubber ducky session so you also open the calendar. Some of the steps are more isolated, some are relying on bigger context - context from multiple systems. Multiple Systems you have to open up, navigate through, fill out different input fields, select different options…

That is where I think agents become interesting: one interaction layer that can sit across tools.

A web UI is very good at showing state.
It can show dashboards, timelines, details, warnings, diffs, logs, metrics, permissions, and the current shape of the system. When I want to understand what is going on, I usually want a good UI. I want to look around. I want to inspect. Good (and clean) visualisation is key for me.

But creating or changing data is often different. My attention may not be on the application that owns the data. It may be in a codebase, a planning document, a conversation, a terminal, a calendar, or an assistant that already has the relevant context. That is also why “just add a chat box” is not an interesting use case of AI for me. You still have to change user interfaces. The fact they are both natural language does not yet do it for me.

I use OpenClaw, and enjoy putting in some work to build myself assistance.
An assistance that knows about the projects I’m involved in. Some of them I manage fully locally, some of them are GitLab projects, where we work with issues (well, mostly). Understanding about milestone progress, new issues, label changes, etc. in just two different projects is a lot of clicking. Or a bunch of API calls neatly orchestrated by a script - managed by AI.

A web UI has to support many use cases, accommodate different UX preferences, and still look good. That is hard to build. Being able to get the information you want by just describing what you want is a lot easier. And you can now integrate multiple systems just like that.

The biggest change from using AI is:
When I build tools for myself, I think about what I really want, rather than about abstraction. Specific flows that I actually use, rather than lots of options making it complex. It helps a lot that my assistant can just build the tools themselves. (I still want a proper web UI for everything visual though, that often looks a bit rough in terminal ;).)

Give me a good API to build my data manipulation requests against, and give me a nice web UI to show what the data truly tells. And show it with intent, not “oh no, how do we make this editable?”.

So yeah, next time you build the complex input formular, maybe just build a good API and leverage natural language interaction with intent.