Webwright
About
Webwright, a Microsoft Research project, replaces the conventional web-agent loop of predicting one click or keystroke in a persistent browser with code-as-action: the model writes and debugs Python Playwright scripts in a terminal, treating each browser as disposable and the local workspace of code, logs, and screenshots as the durable state. The harness is deliberately minimal — a roughly 450-line core loop with httpx and pydantic as the only core dependencies — and it ships plugin manifests for Claude Code and OpenAI Codex plus skills for OpenClaw and Hermes Agent, all sharing one skills folder. On Online-Mind2Web it reaches 86.7% with GPT-5.4, and on the long-horizon Odysseys benchmark it set a reported SOTA of around 60% with GPT-5.4. Its Skill Factory distills solved tasks into parameterized CLI tools that rerun with zero tokens, lifting WebArena accuracy from 55% to 70%. It is aimed at researchers building browser agents on top of coding models.