MCP Playwright and the Subtleties of UI Test Dramaturgy with AI

LLM agents are “blind” to the actual browser, while raw DOM data overwhelms the context and is unstable. The solution lies in an architectural shift toward the Accessibility Tree (achieving a 10x reduction in size) and standardised control via MCP Playwright.

The talk covers:

MCP tool categories (navigation, interaction, data, sessions), integration with SberBrowser, and escaping issues in GigaIDE. We will present live use cases: creating a temporary email address on fex.plus and a “test-fixer” agent that repairs locators based on error messages. We will discuss MCP pain points—specifically token consumption and LLM non-determinism in CI/CD pipelines—and a hybrid MCP + CLI approach as a way forward. Finally, we will demonstrate multi-agent networks featuring an orchestrator and self-healing capabilities driven by multiple