What is Browser agent?

There is no single formal specification for browser agents. The page cited here is Microsoft's README for Playwright MCP, which describes one implementation, so it shows one design rather than a standard.

Playwright MCP is a server that follows the Model Context Protocol and lets a language model automate a web browser through Playwright. It exposes browser actions such as navigating, clicking, typing and taking snapshots as tools that an MCP client can call.

The README highlights one design choice. Instead of sending screenshots for the model to interpret as pixels, the server gives the model a structured accessibility snapshot of the page. The model reads page elements as structured text and refers to them by reference. According to the page, this means the approach does not need a vision-capable model and avoids some ambiguity found in screenshot-driven automation.

This contrasts with screenshot-based approaches, where the model has to interpret images of the screen. The README notes that the server works with MCP clients such as VS Code, Claude Desktop and Cursor, and that Node.js 18 or newer is required.