I gave four coding agents $100 budget to build a PDF editor (blog.nielstron.de)

A recent experiment tasked four AI coding agents with a $100 budget to develop a functional PDF editor. Despite achieving basic functionality, the implementations fell short of usability and often contained bugs, underscoring the limitations of current AI systems in software development. The agents, which include models like Gemini 3.8 Flash and Opus 5, failed to create native applications or intuitive user interfaces, demonstrating a disconnect from human-centric design. For instance, Gemini's attempt at inserting images resulted in poor usability, while Astra's interface loaded buttons with unnecessary textual descriptions, contradicting established design principles. This experiment highlights a pressing issue in AI/ML: agents currently lack the human-like interaction capabilities necessary for recognizing and rectifying usability concerns. While coding agents have advanced in translating natural language into code, their inability to genuinely "experience" software interactions leads to persistent problems, making them reliant on human oversight for practical application. As the tech community looks toward a future where AI automates more of the software development process, this discrepancy emphasizes the ongoing need for human developers, particularly in areas like user experience and interface design. The findings provoke important questions about the viability of fully automated software creation and the enduring role of humans in the development loop.