Agentic Automation: Developing an Autonomous Agent for Automated Testing and Quality Assessment ofWeb Applications

Abstract

=== The growing complexity of modern web interfaces has made conventional functional testing brittle and costly to maintain, as it relies on fixed selectors and is vulnerable to visual or structural changes. In this context, this dissertation proposes the development of an Autonomous Web Testing Agent capable of analyzing, perceiving, and interacting with real web interfaces in a self-directed, efficient, and adaptive manner. The proposed solution is built around an iterative perceive-reason-act-check cycle, implemented in Python using LangGraph for agent loop orchestration, without dependency on commercial test automation platforms. The agent combines multimodal perception by integrating information from the DOM and screenshots with reasoning based on language and vision models (LLM/VLM) to generate multi-action plans. At each agent iteration, the system validates and executes structured action sequences, invoking advanced self-healing mechanisms whenever an execution failure occurs. The self-healing module explores additional recovery approaches, including attribute-based recovery, semantic matching, coordinate-based interaction, and VLM-guided visual localization, enabling the agent to adapt automatically to unforeseen changes in the web interface. Three quantitative metrics for interface evaluation are also implemented: a semantic quality measure (SFS), a structural testability measure (VRC), and a markup compression coefficient (MCC): 1. Semantic Fidelity Score (SFS); 2. Visual Relational Coherence (VRC); 3. Markup Compression Coefficient (MCC). These metrics enable an objective characterization of the structural, semantic, and visual integrity of the tested pages. The solution is evaluated on a set of reference web applications, and additionally on publicly accessible external applications, demonstrating the agent's ability to generalize to real-world contexts. The results show that the agent identifies real issues spanning accessibility, form validation, visual coherence, and security in arbitrary web applications without application-specific configuration. On the benchmark with four synthetic web applications, the tool achieved a mean issue recall of 0.867 and mean F1 of 0.844, outperforming the axe-core baseline by an average margin of 0.75 recall points. In the qualitative evaluation on external applications, the agent identified 23 issues on SauceDemo (versus 3 by axe-core) and 53 on the Zeus application, spanning all five evaluated quality dimensions. The self-healing mechanism was triggered on 38% of actions executed on Zeus, demonstrating its utility in production interfaces with dynamic selectors. Keywords: Autonomous Agents, Web Testing, LLM, VLM, Computer Vision, Multimodal Perception, Self-Healing, Intelligent Automation