Every browser automation you've ever built ran on Chromium. A browser designed for humans, rendering pixels nobody looks at, burning compute on a screen no one sees. Cloudflare shipped Kitesurf earlier this month. A browser built from scratch for agents, with none of that. The number: 3 to 7 times less CPU and memory than Chromium on common agent tasks like screenshots and pulling HTML off a page. Same work, a fraction of the compute. The part that actually matters for you: it speaks the same protocol as what you already use. Puppeteer, Playwright, MCP clients. Your existing scripts work. And it's free while it's in beta. Where this pays off. 1. Anything you're currently paying to scrape. If you're running headless Chrome on a server or paying a scraping service by the request, this is a straight cost cut on the same job. 2. Bursty workloads. It's stateless and runs on Cloudflare's edge, so you can spin up a lot of sessions at once instead of queuing them behind one browser instance. 3. Research and enrichment agents. Screenshots, HTML extraction, PDF generation. The boring high-volume stuff where compute cost compounds quietly. 4. Any tool your client uses that has no API. Same principle as browser control anywhere else. If a human can click it, an agent can. This just makes running that at volume cheaper. Two honest notes, because Cloudflare says both of these themselves. It is not Chromium and it will not render every site. It passes over 235,000 standards tests, but Cloudflare is explicit that passing standards tests is not the same as working on every real website. Test your specific targets before you migrate anything. And it's beta. Free right now, priced later. Don't build a client deliverable on a free tier that hasn't announced what it costs. For a decade we pointed automation at a browser built for people and paid for the parts we never used. That's the thing that just changed.

Microsoft released a tool last week called ThinkingBox. Open source, free, and it should make everyone building client automations a little uncomfortable. Here's what it does differently. Most agent testing reads the transcript. You look at what the agent said it did, and it sounds right, so you ship it. ThinkingBox ignores the transcript completely and checks the database. What actually changed. What records actually got written. They ran 12 models across 507 business tasks in five domains, retail, travel, insurance, banking support, and IT support. Twenty separate trials each, so the same task over and over. This is the finding that matters. Most of the failures didn't look like failures. No crash. No error message. The agent took what looked like valid actions, finished cleanly, and quietly did the wrong thing. Or did part of the right thing and stopped. Think about what that means for a client automation. It's not going to page you. Nobody gets an alert. The agent just runs for three weeks doing 70 percent of the job while everyone assumes it's fine. Here's what to do about it. 1. Stop testing your agents once. One successful run proves nothing. Run the same task twenty times and count how many actually completed. That number is your real reliability, and it will be lower than you think. 2. Check the outcome, not the output. Don't read what the agent reported. Query the CRM. Look at the record. Did the field actually change? Did the email actually send? That's the only test that counts. 3. Add assertions to every client workflow you run. After the agent finishes, something should verify the end state and scream if it's wrong. Silence is not success. 4. Audit what's already live. If you have agents running in client accounts right now, go check the data they've been writing. Not the logs. The data. One honest note. This doesn't mean agents don't work. It means the way most people test them cannot tell the difference between working and looking like it worked. Those are very different things when it's a client's revenue.