I've spent years turning small annoyances into software projects.
Something doesn't work quite the way I'd like? Build an app. Don't want to pay another $15/month subscription? Surely I can make my own version.
With AI, this habit has become almost an obsession.
I can go from an idea to working software in an afternoon. As someone who's spent much of their career building products, I'm still thrilled and amazed by that.
But lately I've started noticing something about how I actually use AI, and it makes me question whether more software is really what I want.
Exhibit A: The agents that stuck around
Over the past year, I've experimented with more AI agents than I can count. Different frameworks, different models, increasingly elaborate setups. Plenty seemed useful until I started living with them.
Most didn't stick.
Two that have are remarkably simple.
I have an always-on Mac mini and Linux machine, each running a long-lived Claude Code session with access to my projects. I can reach them from my laptop or phone, tell them what I need, and have them do the work on the machine where everything already lives.
There's nothing revolutionary about the setup, but I use it constantly.
They're essentially replacing screen sharing and remote desktop.
Turns out, most of the time I don't need to see what's happening on another computer. I just want something done on it.
So despite all the hype around autonomous agents, perhaps another agent isn't what I need. I just need a better way to use the computers I already own.
Exhibit B: I don't want to use your interface
The more I use AI this way, the more I notice how much time I spend operating software rather than accomplishing what I opened it to do.
Need something from a SaaS dashboard? Find the right report, adjust the date range, set a few filters, and export a CSV.
Need to update something in a CRM? First, remember which of its seventeen screens contains the field you're looking for.
None of this is especially difficult. It's just tedious.
And honestly, some genuinely useful products have truly awful interfaces.
I say this as someone who's spent a good chunk of his career designing them.
There's a bit of irony here. With eye.ai, we've built an MCP integration and a Figma plugin that let people analyze designs directly in the tools they're already using. There's no separate analysis interface to open.
Which is a little funny considering how much time we've spent designing exactly that kind of interface at EyeQuant.
And I think that's great.
I still need the underlying products, their data, integrations, infrastructure, and business logic. I just don't always want to use their interfaces.
We've spent decades supposedly making software easier for humans to use. It's quite amusing that one of AI's most useful applications might be saving us from having to use those same interfaces we've been slaving away on.
Of course, sometimes a visual interface is exactly what I want. Especially when I'm exploring, designing, or trying to understand something.
But when I already know what I need done? I'd happily skip the clicking.
Exhibit C: The problem with building everything
Here's where things get a little ridiculous.
While I'm discovering that I want to interact with less software, AI has made it incredibly easy for me to create more of it.
And I've taken full advantage.
I've been building Dapple, my own iOS browser, and Draftwell, where the useful things I create with AI go so I can actually find them again. Both came from having particular ideas about how I wanted those tools to work.
And I genuinely enjoy building them. I use both multiple times a day.
But every new project is also another application that needs updates, maintenance, and fixing when something inevitably breaks.
Some of the smaller tools I've built probably didn't need to exist beyond the afternoon I created them.
Dependencies change. APIs break. Something mysteriously stops working despite nobody touching it.
Sure, AI can help fix those problems too. But it's still another thing that needs looking after.
Basically, I've solved a bunch of problems by giving myself a whole new set of problems.
And the cycle repeats 🥲.
It turns out that being able to build something in an afternoon and wanting to be responsible for it indefinitely are two very different things (who knew!).
Maybe I just want to get something done
I keep coming back to the same realization.
Most of the time, I didn't really want to build an app. I wanted to solve a problem without paying another subscription or dealing with someone else's crappy UI.
And when I use an agent to accomplish something on one of my computers, I don't particularly care that it's an agent. I care that I can get something done without opening a remote desktop session.
The common thread isn't necessarily AI, agents, or even software.
It's the amount of work involved in getting to the outcome I actually want.
I still love building software. And I appreciate a thoughtfully designed interface as much as anyone.
But I'm less convinced that every little inconvenience needs another app, another interface, or another codebase I'll be maintaining five years from now.
I've written before about how the falling cost of building software makes deciding what's worth creating more important.
I'm starting to appreciate that argument a little more every time I open one of my own neglected projects.
AI has made it possible to build software for all my little annoyances. It hasn't made me particularly interested in maintaining software for all of them.
I don't think I want another app. I just want to get something done.