In the realm of automated software testing, the traditional reliance on the Document Object Model (DOM) as the primary method for interacting with web applications is being challenged by the complexities of modern software architectures. Boris Skurikhin, a Co-Founder at Docket, an AI-powered web testing platform, sheds light on the shifting landscape where AI-generated code, complex DOM structures, and the rise of vibe-coded applications are reshaping the foundations of quality assurance in software development.
The prevalent belief that the DOM is the optimal means of web application interaction is now being questioned as software interfaces become more intricate, stitched together from convoluted DOMs and extensive runtime logic. In Silicon Valley, the quality assurance of software has surged in importance, spurred by the proliferation of AI-assisted developers and the repercussions of the LLM boom and vibe-code trend of 2025. What was once considered a specialized craft practiced by skilled engineers has transformed into a routine output from tools like Windsurf, Claude Code, and Lovable.
The integration of AI in software development has undoubtedly brought immense power but also significant responsibility. As AI-generated code becomes increasingly prevalent in platforms crucial for social, financial, and medical needs, the margin for error expands. AI, while a valuable tool, is not infallible, with LLMs exhibiting higher error rates compared to human developers. The deployment of AI-generated code introduces a new layer of complexity to testing, where vibe-coded applications might exhibit functional interfaces on the surface but are underpinned by bloated, disorganized code structures that pose challenges for maintenance and testing.
The conventional reliance on the DOM as a steadfast source of truth in browser automation tools is proving inadequate for today’s AI-generated user interfaces, characterized by minified JavaScript runtimes, canvas-based rendering, and frameworks like Flutter. The evolution of software development has rendered the DOM-centric approach outdated, prompting a shift towards vision-based testing methodologies that operate on what is visually rendered in the browser.
In a world where data undergoes multiple transformations before reaching the end-user through an AI intermediary, the fidelity and granularity of information degrade at each stage, giving rise to what Boris Skurikhin terms the “loss chain.” This gradual loss of detail results in automation tools operating on abstractions that may no longer accurately reflect the on-screen reality, leading to inefficiencies and inaccuracies in testing outcomes.
The limitations of traditional testing approaches have catalyzed a movement towards vision-based testing, which equips automation tools with the capability to perceive and interact with web interfaces visually. By enabling AI agents to interact with elements based on visual cues rather than abstract DOM structures, vision-based testing circumvents the discrepancies introduced by the loss chain, offering a more accurate and reliable testing environment for modern software applications.
Boris Skurikhin’s advocacy for computer vision in automation tools underscores the importance of aligning testing methodologies with the visual nature of human-computer interactions. In a landscape dominated by vibe-coded applications and intricate UI designs, the rendered screen emerges as the ultimate source of truth, superseding the conventional reliance on DOM structures for testing.
While vision-based testing may currently entail higher costs and slower processes compared to DOM-centric frameworks, the rapid advancements in vision models are narrowing this gap. Vision-based approaches excel in scenarios requiring extensive interaction with web interfaces, such as canvases, iframes, and nonstandard input components, making them ideal for tasks involving visual elements and dynamic user interactions.
In conclusion, the evolution of software testing methodologies reflects a paradigm shift towards vision-based approaches that leverage the visual fidelity of web interfaces to enhance automation testing accuracy and efficiency. As the software landscape continues to evolve with the advent of vibe-code and AI-generated applications, embracing vision-based testing represents a proactive step towards ensuring the robustness and reliability of modern software systems.
Takeaways:
– The traditional reliance on the Document Object Model (DOM) in automated software testing is being supplanted by vision-based approaches that leverage computer vision for enhanced accuracy.
– The integration of AI in software development necessitates a shift towards vision-based testing methodologies to align with the complexities of modern AI-generated user interfaces.
– Vision-based testing excels in scenarios requiring extensive interaction with web interfaces, offering superior accuracy in handling visual elements and dynamic user interactions.
– Embracing vision-based testing is essential for adapting to the evolving landscape of software development, characterized by vibe-coded applications and intricate UI designs.
Tags: automation
Read more on forbes.com