End-to-end (E2E) testing is a persistent industrial bottleneck: Selenium, a widely used open-source framework for automating web browsers via scripted interactions that emulate user behaviour, relies on brittle scripts, while off-the-shelf Large Language Models (LLMs) hallucinate selectors and saturate the context window when asked to synthesise complete user journeys over production DOMs. We present Doc2Test, a Multi-Agent System (MAS) deployed inside Wideverse (a software engineering and digital transformation ICT company) that translates unstructured requirement documents into executable Selenium suites through three specialised phases: requirements analysis, dynamic DOM/visual context perception, and adaptive code generation with a self-healing loop. Across 90 independent runs on a production Angular cruise-booking platform (∼145k LoC, under NDA), four open-sourced internal Wideverse applications, and one independently-developed external application (RealWorld), Doc2Test achieves a 94.4% generation success rate (95% CI: [87.5%, 98.2%]), detects 98.9% of 180 injected UI regressions, and reduces token usage by 86% against a zero-shot single-prompt baseline. Over three months of nightly Jenkins operation, it reduced maintenance effort by ∼2.6 person-days / sprint and the flaky-test rate from 11.4% to 4.6%. We further report cost breakdowns and deployment lessons from integrating Doc2Test into our industrial partner’s CI/CD pipeline.

@inproceedings{paperbot12595,
  title = {Doc2Test: An Industrial Multi-agent Framework for Document-Driven End-to-End Selenium Test Generation},
  author = {Andrea Lops and Antonio Ferrara and Fedelucio Narducci and Azzurra Ragone and Michelantonio Trizio},
  booktitle = {ASE},
  year = {2026},
  doi = {10.1145/3832783.3834507}
}