Webwright ของ Microsoft: ให้ AI เขียนโค้ดท่องเว็บ แทนที่จะนั่งคลิกทีละครั้ง

Microsoft's Webwright: AI Writing Web Code Instead of Clicking Step-by-Step

A Microsoft Research study reveals that the issue with AI agents working on the web is not just about occasional clicking errors, but rather the iterative nature of the agent's workflow. To address this, Microsoft proposes a solution called Webwright, which aims to separate the agent from the browser. This new approach, according to the developers, offers several advantages.

Unlike traditional browser agents, Webwright allows the agent to work with a single session of the browser, while the agent receives the current state of the browser session and predicts the next action. This prediction can be a click, typing, hovering, selecting an element, or invoking a short web tool. The agent is constrained to make predictions one step at a time, within a specified loop.

After completing the work, Webwright leaves behind a single sequence of clicks made during a single execution, rather than a reusable action. The idea is to separate the agent from the browser, treating the browser as a tool that the agent uses to open, inspect, and discard as needed during the software development process. This approach simplifies the system, with minimal components: a terminal, a browser, and the model itself.