What is Playwright

Playwright is Microsoft’s open-source framework for browser automation and end-to-end testing. On its own it’s a testing tool; paired with Playwright MCP, it’s also become the default way an AI agent gets hands on a real web browser.

---

As a testing framework

Feature	What it means in practice
Multi-browser	One API drives Chromium, Firefox, and WebKit — cross-browser coverage without three separate test suites
Auto-wait	Actions wait for elements to be actionable (visible, enabled, stable) before firing — eliminates most flaky-test sleep() hacks
Trace viewer	Records a full timeline (DOM snapshots, network, console) for any test run — post-mortem debugging without re-running
Codegen	Records your manual clicks/typing and emits the equivalent test script
Isolated contexts	Each test gets its own browser context (cookies, storage) — no cross-test state leakage

npx playwright test              # run the suite
npx playwright test --trace on   # capture full trace for debugging
npx playwright codegen <url>     # record actions → generate script

As an agent tool: Playwright MCP

The reason Playwright shows up on an AI-agent pipeline backlog rather than just a QA backlog: microsoft/playwright-mcp exposes Playwright as an MCP server, so any MCP-aware agent (Claude, an ADK agent, a CrewAI tool) can drive a real browser.

The key design choice is accessibility snapshots, not screenshots:

Traditional "computer use":  screenshot → vision model guesses coordinates → click(x, y)
Playwright MCP:              accessibility tree → deterministic element refs → click(ref)

That difference matters operationally:

• No vision model required — works with any text-only LLM
• Actions target a stable element reference, not a pixel coordinate — far less brittle when layout shifts
• Every action is inspectable/auditable against the accessibility tree, which is what makes this safe enough to hand to an autonomous agent in the first place

LLM ──▶ Playwright MCP server ──▶ real browser (Chromium/Firefox/WebKit)
                │
                └─ navigate / click / type / fill form / screenshot / read accessibility tree

By 2026 this pattern has effectively become the standard: several competing Playwright-based MCP servers exist, but Microsoft’s reference implementation is the default most agent frameworks wire up first, with alternatives differentiating mainly on token efficiency.

Why it’s on the backlog: it’s the concrete answer to “how does my agent actually interact with a UI that has no API” — relevant both for browser-based tool use in What is Hermes Agent / What is CrewAI-style pipelines, and for testing whatever agent pipeline gets built.