Your agent needs browser automation that actually works

Most AI agents live in text bubbles. They can write code, analyze data, and chat endlessly — but ask them to fill out a form, click through a checkout flow, or test your actual app? They're useless.

Browser automation changes this. But most solutions are either toy demos (Playwright scripts that break when a button moves) or enterprise monsters (Selenium farms that need a DevOps team).

The Vercel team just shipped something different: agent-browser, a CLI that gives AI agents real browser control without the usual pain.

Here's what makes it work: it runs headless by default, handles viewport management automatically, and gives your agent structured ways to interact with pages — not just "click the thing that looks clickable."

I've been running it with Claude and Cursor for the past week. The setup is dead simple:

npm install -g @vercel/agent-browser
agent-browser --port 3000

Now your agent can navigate to pages, fill forms, take screenshots, and extract data. But here's the key insight: don't give it free rein to "browse the web." Give it specific missions.

Mission-based browser automation works. Exploration doesn't.

Instead of "research our competitors," try:

• "Go to competitor.com/pricing, screenshot their plans, extract the feature list from their Pro tier"
• "Fill out our demo form with test data, screenshot any validation errors"
• "Navigate our checkout flow, stop at payment, report any UX friction"

The agent-browser CLI gives you structured outputs for each action. Your agent gets JSON responses with element selectors, page state, and error details. This means it can actually debug what went wrong instead of just saying "the page didn't work."

I'm using it for three things that used to require manual work:

1. Automated user testing. My agent runs through our signup flow every morning, screenshots any broken states, and reports conversion funnel drops.

2. Competitor monitoring. It checks pricing pages, feature announcements, and job postings. Structured data extraction means I get clean reports, not messy screenshots.

3. Form automation. Directory submissions, partnership applications, feedback forms — anything that's too tedious for humans but too nuanced for basic HTTP requests.

The real power shows up when you combine it with other agent skills. My agent browses to a competitor's blog, extracts their recent posts, analyzes the content strategy, then drafts our response posts. End-to-end automation that actually ships results.

Warning: Browser automation can burn through tokens fast. Set clear boundaries on how many pages your agent can visit per session. I use a 10-page limit for exploration tasks, unlimited for specific workflows.

The CLI works with any agent that can run terminal commands — Claude Code, Cursor, Gemini CLI, or your custom setup. No special integrations needed.

If you're running production agents that need to interact with the real web (not just APIs), browser automation isn't optional anymore. It's the difference between an agent that talks about work and one that actually does it.