Playwright new Test Agents explained (bug0.com)

Playwright introduced "Test Agents" — an AI-driven trio (Planner, Generator, Healer) that plan, generate, and automatically repair end-to-end browser tests. The Planner explores an app and emits Markdown test plans, the Generator converts those plans into runnable Playwright test files (verifying selectors and assertions), and the Healer monitors failures, replays failing steps, suggests patches and re-runs until tests pass. Agents live in a repo layout (.github/ for agent defs, specs/ for plans, tests/ for generated code) and can be bootstrapped with npx playwright init-agents --loop=<env>. Under the hood the system combines Playwright’s browser engine, an LLM layer (GPT/Claude-like models) and an orchestration loop that exchanges structured commands over the open-source Model Context Protocol (MCP), enabling auditable, safe JSON-style interactions (e.g., getElements, click) rather than arbitrary code execution. For the AI/ML and dev-tooling community this is significant: it automates repetitive QA work, accelerates coverage growth, and introduces a standardized, model-agnostic bridge (MCP) for LLMs to safely interact with developer tooling. Limitations remain — reliance on stable locators, reactive (not proactive) healing, and model variance — but Test Agents mark a practical step toward intent-based testing (natural-language goals driving resilient tests). The move also catalyzes an ecosystem of hybrid tools (Reflect, Stagehand, Bug0, etc.) exploring complementary approaches to autonomous, maintainable browser testing.