Touring_test: A Cucumber Extension for Agentic Usability Testing (worksonmymachine.ai)

touring_test is a new Cucumber extension (a Ruby gem) that replaces DOM-driven test scripts with agentic “usability” tests: it sends a natural-language instruction plus a screenshot to Gemini’s computer-use model, which then decides and executes actions (click, type, scroll, navigate) via tool calls, captures follow-up screenshots, and iterates until it succeeds or gives up. The project reframes BDD for the age of agents — “Agentic BDD” (ABDD) — by testing whether a reasonable, confused entity can discover and complete tasks (e.g., “log out”) instead of asserting the existence of a specific DOM element or selector. The code is open on GitHub under the new org Works On Your Machine and currently supports Gemini’s computer-use capability. Technically, touring_test gives the agent only pixels and the instruction — no DOM, no CSS, no element IDs — so failures are often user-like (misclicks, misinterpretations, hallucinated success). That makes tests noisier, slower, and costlier, but potentially far more valuable for catching UX and discoverability problems that Selenium-style tests miss. The pattern also suggests a feedback loop: let agents attempt tasks, surface human-like failures, and feed those failures to coding agents or developers to iterate on UX. It’s an experiment in automating “confusion” rather than superhuman precision — and it may expose the gaps between automated assurance and real user experience.