ianalloway's picture
Upload folder using huggingface_hub
d3ebf05 verified
|
Raw
History Blame Contribute Delete
1.57 kB
metadata
title: Browser Agent Benchmark
emoji: 🧭
colorFrom: blue
colorTo: purple
sdk: static
pinned: false
license: mit
short_description: Real-world browser automation as an agent benchmark

The browser is the real agent benchmark

Most agent demos skip the part that breaks in production: real websites.

This Space is a small public companion to Ian Alloway's write-up on real-world browser agents, publishing automation, and why "I clicked the button" is not enough unless the system leaves a receipt.

Read the full post: https://allowayai.substack.com/p/the-browser-is-the-real-agent-benchmark

What this benchmark cares about

A useful browser agent should survive the things modern web apps actually do:

  • authenticated persistent browser profiles;
  • MFA / CAPTCHA / human handoff without leaking credentials;
  • iframe and Shadow DOM traversal;
  • rich-text editors like ProseMirror and Quill;
  • native passkey/WebAuthn prompts that sit outside the DOM;
  • final-state verification instead of trusting a successful process exit;
  • locked-down remote desktop exposure after manual auth.

Receipts over vibes

The bar is simple: if the agent claims it posted, submitted, updated, or paid for something, it should produce a concrete receipt: a final URL, a DOM state, a server response, a public page check, or a screenshot.

That is the difference between a chatbot and infrastructure.

Links