In the experiment, the researcher provided Gemini 3.8 Flash, GPT Astra 6, Opus 5, and Fable 5 each with a $100 budget and tasked them with creating an open-source PDF editor that would have a great user experience. The agents were instructed to inspect existing tools and search for user-reported shortcomings before building. Despite these instructions, all four agents produced implementations that the researcher described as functional but not great, noting that bugs appeared within just a few clicks of use.
The agents’ outputs showed several recurring shortcomings. Most built web applications rather than native installable programs, which limits their usefulness as PDF editors that should ideally be available system-wide. Interface design choices were problematic: one agent added literal text descriptions to every button in an attempt at intuitiveness, while another created a profusion of icons with unclear meaning, and none included a search or help feature for settings. Performance suffered from inefficient backend processing, with sequential requests taking seconds and mixed SQL queries combined with manual joins in Python. Ambitious features such as editing existing PDF text, OCR, or vector graphic manipulation were largely missing, with only Opus attempting text editing and only Gemini implementing compression. Platform support was weak, as touch scrolling was jagged in Fable’s app and Gemini’s layout did not adapt to mobile screens, leaving controls half‑visible.
The researcher attributes these issues to a fundamental difference in how AI agents and humans interact with software during development. Humans work with concrete goals and notice friction when they encounter confusing interfaces or slow operations, whereas the agents reportedly do not pursue concrete goals, experience less friction, and examine the UI far less often. This disconnect leads to missed visual bugs—such as darkened‑mode color inversion or skewed exported images—and a failure to optimize performance bottlenecks that impede real‑world use. Without changes to how large language models operate and interact with development environments, the author warns that AI coding agents may continue to produce “SlopWare”: software that is technically functional but falls short of the polish and usability expected in professional tools.