Microsoft Research unveils Webwright: AI agents write web-navigation code instead of clicking
Microsoft Research introduced Webwright, an approach where an AI agent gets a terminal and writes programs to navigate the web instead of clicking step by step. On the long-horizon Odysseys benchmark, GPT-5.4 jumped from 33.5% to 60.1%, and reached 86.7% on Online-Mind2Web.
- Webwright gives the agent a terminal and Playwright on Chromium instead of DOM clicks
- On Odysseys, GPT-5.4 rose from 33.5% to 60.1% at 76.1 steps per task
- Online-Mind2Web score is 86.7%, versus 84.7% for Claude Opus 4.7
- Cost per task: $2.37 with GPT-5.4 and $6.09 with Claude Opus 4.7
Read next
Software