microsoft/Webwright

What it solves

Webwright addresses the bottleneck of traditional web agents that rely on step-by-step, single-action predictions (like clicking a specific coordinate). These traditional methods are often fragile and struggle with long-horizon tasks. Webwright instead treats the browser as a disposable environment that a coding agent can launch, inspect, and control via Python scripts, making web interactions more robust, reusable, and easier to debug.

How it works

Rather than predicting a single next action, Webwright provides an LLM with a terminal where it can write and execute free-form Python scripts using Playwright. The agent iterates by writing code, executing it, inspecting screenshots and logs, and repairing the code until the task is complete. The state is maintained in the local workspace (code and logs) rather than the browser session itself.

Who it’s for

Developers building browser-using agents or those who want to enhance existing coding agents (like Claude Code or OpenAI Codex) with professional web automation capabilities.

Highlights

• Code-as-Action: Uses Python scripts instead of discrete tool calls, allowing for loops, functions, and complex logic.
• Skill Factory: Distills successful task solves into reusable, parameterized Python skills that can be run standalone without an LLM, significantly increasing accuracy and reducing token costs.
• Lightweight Architecture: A minimal core loop with zero hidden frameworks, making it easy to debug and fork.
• Plugin Support: Ready-to-use manifests for integration into Claude Code, OpenAI Codex, OpenClaw, and Hermes Agent.
• High Performance: Achieves state-of-the-art results on Online-Mind2Web and Odysseys benchmarks for long-horizon web tasks.