Skip to content

Webwright: A SWE-Style Browser Agent Framework

Source: Microsoft/Webwright on GitHub
Date Published: 2026-05-26
Author: Microsoft Research

TL;DR

Webwright is a radically simple browser agent framework that achieves state-of-the-art results on long-horizon web tasks. Its key insight: separate the agent from the browser, treating the browser as a disposable environment the agent spawns while developing a program. No multi-agent systems, no graph engines, no plugin layers — just a terminal, a browser, and a model.

Core Philosophy

"Webwright gives LLM a terminal where it can launch multiple browser sessions to inspect the page and complete a web task."

"Separate the agent from the browser, and treat the browser as something the agent can launch, inspect, and discard while developing a program. The persistent artifact is not the browser session — it's the code and logs in the local workspace."

Architecture Comparison

Stagehand browser-use Webwright
Paradigm Hybrid: code + NL primitives Autonomous LLM loop over DOM/AX Coding agent with terminal
Action Space Playwright code / NL Indexed click/type actions Free-form Python (Playwright scripts)
State Browser session Browser session Local workspace (code, screenshots, logs)
Loop shape Imperative observe→predict→execute→repeat write code → execute → inspect → repair

Key Dependencies (~1.5k LoC)

httpx
pydantic
playwright
typer

SOTA Performance

Achieved state-of-the-art on two real-website benchmarks with a 100-step budget:

  • Odysseys (long-horizon evaluation)
  • Online-Mind2Web (AutoEval)

Project Structure

webwright/
├── pyproject.toml
├── src/webwright/
│   ├── run/cli.py                  # CLI entrypoint
│   ├── agents/default.py           # Core agent loop
│   ├── environments/               # Playwright browser workspace
│   ├── tools/                      # image_qa, self_reflection
│   ├── models/                     # OpenAI/Anthropic backends
│   └── config/                     # Stackable YAML configs
├── tests/
└── outputs/                        # Run artifacts

Quick Start

pip install webwright
webwright run -t "Your task instruction" --start-url "https://example.com"

Key Takeaways

  1. Webwright achieves SOTA by treating the browser as a disposable environment, not a persistent session
  2. The agent writes and executes code rather than predicting individual web actions
  3. Ultra-minimal codebase (~1.5k LoC) with no complex orchestration layers
  4. Ships as a plugin/skill for existing coding agents (Claude Code, Codex CLI)
  5. The "code-as-action" paradigm represents a fundamentally different approach from traditional browser agents