@shao__meng: Microsoft releases terminal-native Web Agent framework: Webwright https://github.com/microsoft/webwright… Core design: code as action. Traditional web agents follow a loop of 'observe → predict next click → execute', where each step relies on L…

X AI KOLs Timeline Tools
web-agent microsoft open-source playwright terminal automation framework

Summary

Microsoft has released a terminal-native Web Agent framework called Webwright, which automates web operations by having LLMs write Playwright scripts. It features a minimal architecture and state-of-the-art performance, with support for multiple model backends and product integrations.

Microsoft releases terminal-native Web Agent framework: Webwright https://github.com/microsoft/webwright… Core design: Code as action Traditional web agents follow a loop of "observe → predict next click → execute", where each step relies on LLM judgment. Webwright's approach is closer to a software engineer's mindset: · Let the LLM write Playwright scripts — turning web operations into runnable Python programs · Workspace as state — scripts, screenshots, logs saved locally, browser sessions can be rebuilt anytime · Terminal-first — the core loop consists of only three modules totaling about 1000 lines of code, with no hidden orchestration layer The "byproduct" of this pattern is reusable automation programs, not one-time interaction traces. Performance: SOTA level · Online-Mind2Web (300 tasks): 86.7% (GPT-5.4), highest among comparable open-source frameworks · Odysseys (200 long-horizon tasks, average 76.1 steps): 60.1% (GPT-5.4), +15.6pt over previous SOTA, +26.6pt over baseline GPT-5.4 · Claude Opus 4.7: 84.7% / hard cases 80.5%, surpassing GPT-5.4 on hard cases Minimal architecture Runner (150 lines) -> Model Endpoint (550 lines) -> Environment (300 lines) · Only depends on httpx, pydantic, playwright, typer · No multi-agent system, no graph engine, no plugin layer · Supports OpenAI, Anthropic, OpenRouter backends Productization and integration · Claude Code: plugin /plugin install webwright@webwright, supports /webwright:run and /webwright:craft · OpenAI Codex: install from plugin marketplace, invoke via @ webwright · OpenClaw / Hermes: share skills/webwright/ directory, unified skill specification Key innovations · Task2UI mode (added 2026-05-11) — automatically renders results as an HTML app after task completion, making outputs visualizable and reusable · Script reusability — even a small model like Qwen-3.5-9B can achieve 66.2% hard-case completion rate with the help of pre-built tool scripts · Auditability — every run saves traces, screenshots, and reports for easy debugging and regression
Original Article
View Cached Full Text

Cached at: 05/27/26, 09:21 AM

Turn Your Coding Models to Be State-of-the-art Browser Agents

Similar Articles

@axichuhai: This Alibaba open-source project, Page-Agent, allows you to control web interfaces using natural language. It has already garnered 18.7K stars on GitHub. It injects an AI agent directly into web pages, and you can use natural language to instruct it to click buttons, fill out forms, and navigate workflows. It doesn't need a headless browser, screenshots, OCR, or multimodal models.

X AI KOLs Timeline

Alibaba's open-source project, Page-Agent, lets you directly control web interfaces with natural language, with no need for headless browsers or multimodal models. It has earned 18.7K stars on GitHub.

@CycleDecoded: WebRover is an open-source AI agent. Give it a sentence, and it opens the browser, identifies webpage elements, clicks, turns pages, scrapes data, gets the job done, and finally organizes the results you want clearly. License: MIT License. Positioning: Autonomous web automation AI agent…

X AI KOLs Timeline

WebRover is an MIT-licensed open-source AI agent that uses natural language to drive the browser for web automation, cross-site data scraping, and deep research, with support for local deployment.

@CycleDecoded: Programmers and terminal freebie hunters, look here — someone actually built this. In the past, writing code meant opening a bunch of windows, switching back and forth between the black Terminal, VSCode, browser, and ChatGPT until your hands were worn out. The open-source programmer project Wave Terminal (waveterm) directly combines the command line, ...

X AI KOLs Timeline

Introducing the open-source AI terminal Wave Terminal, which integrates the command line, editor, browser, and AI into one terminal, with support for inline previews, persistent SSH sessions, and multi-model access. It has received 22k+ Stars on GitHub.