Ephemeral Software

This week, I built an AI agent that can modify its own code while the app is running, and it is completely changing how I think about software.

This happened last night: a non-technical friend needed a recipe book. In 15 minutes, an AI agent built him a working app. But here's where it gets wild: when he wanted to add his first recipe, he just texted the agent a URL. The agent scraped the site, parsed the recipe, stored it via API, and updated the UI. No feature request. No deployment cycle. The app evolved in real-time from a text message. There isn't even a dedicated box to enter the URL, no custom logic exists anywhere. It was all done on the fly.

Another non-technical friend wanted to upload some data into the app they were building so they just asked the AI to add the functionality to their app and within a few minutes they were unblocked.

I haven't been able to stop building since. Home Assistant interfaces, Tailscale dashboards, Proxmox UIs, a trip planner, apps for my kids. Each took minutes, not months. More importantly, they're alive. Text "make that button blue" or "add a calendar view" and watch the app reshape itself instantly. No bug queues, no prioritization meetings, no status updates. Just immediate transformation.

See, all of this was possible before, but it was so time consuming and expensive that it wasn't worth doing. I would have had to find time and actually do the work instead of any of the other billion things I keep putting off.

Since personal computers arrived, we've consumed software built by others. We learned their interfaces, adapted to their workflows, clicked where they told us to click. This made sense when software development cost millions and took years. But what happens when that cost drops to nearly zero and development time shrinks to minutes?

I believe this is a monumental shift in how we interact with computers. It's crazy enough that we can text with a computer like we text with a human, but now when the computer can respond back and build apps on the fly to help me do the one thing I want to do in the moment. It feels absolutely surreal.

We're entering an era of personal, disposable, just-in-time software. Instead of installing apps, we describe what we need and watch them materialize. These apps obviously aren't replacements for infrastructure or mission-critical systems. They're a new category entirely, software that's meant to be modified, broken, and thrown away.

Think of it this way, it isn't about building production systems or scaling to millions of users. It's about quick, purpose-built tools that solve immediate problems. We already do this, every developer has written throwaway bash scripts without worrying about maintainability or architecture. The difference is that now anyone can do it, just by describing what they need.

One big open question is whether any of this is secure, and honestly, for a self-hosted recipe app or a personal dashboard, maybe it doesn't matter as much as we think. We're trusting the model to produce the right thing while acknowledging the risks of prompt injection. For throwaway tools, that might be a reasonable trade-off. Maybe sandboxing the agent and controlling what APIs/websites it has access to is sufficient.

The future of software might just be a prompt away. What would you text into existence?