As an experiment I provided ZERO technical guidance, I let it pick architecture, programming language, GPU interface, rendering interface, all of it. I only provided feedback from the perspective of a user and only things visible to a user. I.e. "Adjusting the curves spline is too slow. Make it instantaneous", I did not ask it to downsample the preview or move the operation the the GPU.
As a result I have a fully functional, super fast, stable, Linux based 2D image editor, written in Rust and D, GPU accelerated (Vulcan). Startup time is 0.458 sec. Every operation is instantaneous on a sub $1000 laptop with RTX4060.
I needed all the features that I use personally and I did not care about features I don't use or that I don't know about, so everything I needed was:
- typical format support: png, jpeg, bmp, tiff, targa - layer management (new, delete, lock, merge, duplicate); - layer blending (opacity, fill, and the whole panel of blending types - darken, difference, screen, etc); - filters with preview render for live adjustments; - drawing and selection tools complete with properties for each; - history (undo/redo); - complete UI - fully stylized window, window setup presets and flexible panel allocation; - right click, alt click, shift click support for tools, matching key bindings; - open and save a range of formats; - text insertion (vectorized) and ability to rasterize the text; - vector objects and ability to rasterize them; - testing framework that can simulate mouse clicks, drawing and UI interactions; - complete regression testing suite
I dont know Rust, I dont know D, I dont know how to use Vulcan. I know image processing and software engineering but it didnt matter at all. I only provided typical user feedback - i.e. "I cant see layer controls in the UI" or "curves adjustment is lagging when I move the spline, must be instantaneous". Developing this software required no skill, insight or know-how.
I added a few further improvements that I noticed missing so the latest version has more features with a total agentic time at 10 hours and my own time at 3 hours.
I am now using it as a replacement for Photoshop that costs $275/year in a predatory subscription where you cant cancel without a $200 penalty.
The point is that software will eventually become worthless when everyone can just spin their own solution that accurately reflects their needs and skills.
Why pay $275/year in predatory subscription for a bloat of features that you dont use when you can leave the agents overnight and wake up to the product thats perfectly tailored to you and you can change any part of it that you want? This is feasible today and it will become a standard tomorrow.
Software is not a product or service anymore. Its a tool used to solve a problem thats described in English.