Context
browser_agent_mvp is a local browser agent that controls visible Chromium through Playwright. It focuses on practical browser automation without relying on predefined selectors for every action.
Problem
Browser agents need to reason over changing pages, choose actions, remember recent steps, and leave logs that can be inspected. A brittle selector script is not enough for multi-step tasks.
Constraints
- The browser remains visible for inspection.
- Actions are represented as structured tool calls.
- The agent loop needs memory and result logging.
- External connection through MCP tools is part of the design.
Approach
The agent reads an ARIA snapshot of the page, chooses structured actions, executes them through Playwright, stores recent action memory, and logs results for review.
Implementation highlights
- ARIA snapshots as the page representation layer.
- Tool calls described with schemas.
- Recent action memory to reduce repeated or contradictory steps.
- MCP server exposing browser tools for external agent connection.
Result
The project demonstrates practical agent loop design for browser automation, with explicit state, visible execution, and reviewable logs.
What it demonstrates
Agent loop design, Playwright automation, MCP tool exposure, and human-reviewable browser workflows.