Microsoft Research releases Webwright, a browser agent that solves 60% of web tasks

Microsoft Research has unveiled Webwright, a framework for browser agents that completes complex web tasks almost twice as successfully as base language models.

How Webwright works

It is a terminal-based agent that automates browser interaction. The key difference: instead of the familiar click-trace approach (where the system records a sequence of clicks and coordinates), Webwright generates and executes Playwright scripts, a powerful framework for programmatic browser automation.

The framework is simple: roughly 1000 lines of code and three modules running in a single agent loop. The minimalist design looks naive at first, but the results turned out to be impressive. Rather than trying to generate point-by-point clicks, the agent works out the structure of the DOM and writes the scripts it needs.

Benchmark results

On Odysseys (which tests long web tasks in a real browser), Webwright with GPT-5.4 reached 60.1%. That is twice the 33.5% of the base model alone. On the simpler Online-Mind2Web benchmark the score is even higher: 86.7%. Importantly, this is the best result among all open-source harness recipes.

The twofold improvement does not come from special tricks or hardcoded solutions. It is the straightforward result of a well-designed agent loop and effective use of GPT-5.4's capabilities.

• Odysseys benchmark: 60.1% (up from 33.5% for the base model)
• Online-Mind2Web: 86.7% (record among open-source)
• Framework size: ~1000 lines of code
• Architecture: three modules in a single loop
• Model: GPT-5.4 (standard, without fine-tuning)

Why it works

Browser agents have long relied on click-trace sequences or required huge language models. Webwright shows a third path: the right architecture, with Playwright scripts as an intermediate language, delivers a large quality gain.

Playwright also lets the agent work with the DOM directly, which is more reliable than relying on computer vision. When a site changes, the script can adapt because it sees the page structure, not just pixels.

What it means for the market

Browser agents are maturing. Microsoft Research has shown its approach, while OpenAI (Operator), Anthropic (Computer Use) and others are working in parallel. The web automation market is only beginning to form: filling in forms, comparing prices, ordering services, managing subscriptions.

Webwright demonstrates that good results do not require waiting for a super-model: the right architecture and simple modules can deliver a multiple-fold improvement.