Universal Testing Language

title: "Universal Testing Language" slug: "universal-testing-language" summary: "A test-intent DSL that translates to multiple automation frameworks. One specification, many runtimes." categories:

• "Automation Frameworks"
• "AI-Assisted Testing" technologies:
• "TypeScript"
• "Playwright"
• "Python"
• "Selenium"
• "Zod" status: "Concept" order: 1 featured: true

Overview

The Universal Testing Language (UTL) is a domain-specific language for describing the intent of a test once and translating that intent into multiple automation runtimes - Playwright, Selenium, Cypress, and emerging agentic runners. The goal is to decouple what a team wants to verify from the framework it currently happens to use.

Problem

Teams pay a recurring tax for framework lock-in. A migration from Selenium to Playwright, or from one CI to another, often means rewriting test suites - not because the test intent changed, but because the runtime did. AI-generated tests amplify this: when an agent generates a test today, the team is committing to whatever framework the agent chose, forever.

Users

QA architects evaluating long-lived automation strategy; engineering teams considering a framework migration; AI-augmented testing workflows that want a stable abstraction layer underneath generated test cases.

Goals

• Express a test as a structured, framework-neutral description.
• Translate that description into a runnable spec for at least two frameworks.
• Demonstrate that the same UTL document can run on Playwright and Selenium without source changes.
• Be honest about what the abstraction can and cannot do well.

Architecture

The parser produces a typed AST validated with Zod. Each emitter is a pure function from AST to source code. Emitters are independent and swappable - new runtimes plug in by implementing the emitter interface.

Technologies

• TypeScript
• Zod
• Playwright
• Selenium
• Python

Testing Strategy

• Snapshot tests on emitter output keep generated specs stable.
• Each emitter is round-tripped against a sample suite; the emitted spec must actually run and pass against a fixture app.
• AI-generated UTL is checked into the test corpus so regressions across model versions are visible.

AI Role

AI assists at the authoring layer - generating UTL from natural-language intent or from exploratory walks of an application. AI does not run the tests, edit the AST, or modify the emitters. The agent's surface stays above the deterministic runtime, where human review is cheap and the generated artifact is small enough to audit.

Challenges

• The abstraction is leaky when frameworks differ on what is testable (e.g. shadow DOM behavior in Playwright vs Selenium). Emitters must signal "cannot represent" rather than silently degrade.
• Versioning the language matters early - breaking the schema breaks every consumer's tests.

Results

A working concept with two emitters (Playwright + Selenium) and a small sample suite that runs identically on both. Performance overhead from translation is below 5 ms per case.

Next Steps

• Add a Cypress emitter and a contract-test emitter (API-first scenarios).
• Wire an agentic emitter that targets an experimental AI test runner.
• Publish a small CLI so the project can be evaluated by other teams.