Webwright: Why AI Web Agents Should Write Code, Not Click

Many AI web agents are unreliable for long tasks because they operate one fragile click at a time, often failing when a webpage changes unexpectedly. A new approach called Webwright, from Microsoft Research, instead has the AI write a complete program to accomplish the task. This code-writing method significantly boosts success rates on complex web automation, jumping from 33.5% to 60.1% for the same underlying AI model. Crucially, instead of leaving behind a temporary click history, this process creates a reusable command-line tool that can be inspected, modified, and run again.