Squoosh helps brands predict conversion outcomes before changes go live. The team uses fleets of computer use agents, trained on the behavior of real shoppers, to forecast how real users will behave across e-commerce storefronts, self-service portals, and other customer-facing digital experiences. They’re trusted by over 25 leading e-commerce brands and the 2nd largest telecommunications provider in the United Kingdom.
Browser infrastructure sits at the center of how Squoosh works. Each synthetic shopper drives a real browser through the customer’s checkout flow. Customers can open any shopper and watch the exact video recording of that journey, see whether they converted, and understand why. But things weren’t always running so effortlessly.
Problem: Synthetic shoppers that wouldn’t scale and observability that fell short
To get statistically significant test results, Squoosh needs to send enough synthetic shoppers through a customer’s site, which means running a large number of real browsers, reliably, in parallel. They tried to own that infrastructure themselves first, and it was a constant source of pain. Maintaining it pulled the team away from the only thing they wanted to be working on: their models.
“We just wanted to focus fully on making AI models that mimic human shopping behavior. We wanted to give all the other hard technical problems to a partner.”
Tyler Dooskin, Founder, Squoosh
They moved to Browserbase next, and hit a different set of walls. As mentioned, a single test can mean sending thousands of synthetic shoppers through a site, and concurrency limits and cold-start problems capped how many they could run at once, which directly limited the quality and thoroughness of their test results. And because their shoppers need residential proxies to reach customer sites without tripping bot detection, the proxy add-on on Browserbase was significantly expensive, making the cost of every shopper something they had to weigh.
On top of that, video recordings are central to how Squoosh earns trust. Customers want to open any synthetic shopper and watch a real video of that shopper moving through their site, rather than take a row of metrics on faith. That made the quality, durability, and retrieval of session recordings a first-class requirement, not a nice-to-have, and it was another place their earlier infrastructure fell short.
Solution: Selecting Kernel over Browserbase and self-hosting
Squoosh needed managed browser infrastructure that could run thousands of synthetic shoppers in parallel without the concurrency limits and cold starts they hit on Browserbase, with residential proxies and high-quality video recordings included. Tyler built an internal benchmark on performance and speed and ran their self-hosted setup, Browserbase, and Kernel through it. Kernel won decisively.
The benchmark results showed the concurrency and cold-start problems went away, so the scale of a test was no longer capped by the infrastructure underneath it. The built-in proxy mattered just as much: paying for proxies separately had been very expensive on Browserbase, and having residential proxies included with Kernel was a real cost and reliability win. And the video recordings and live view were “rock solid.”
“Kernel blew Browserbase and our own solution out of the water on our internal benchmarks. Plus, the video recordings are rock solid and easy to integrate, and the support team has gotten us over the finish line every time.”
Tyler Dooskin, Founder, Squoosh
Results: 2000x faster A/B tests and thousands of synthetic shoppers per run
Today, Squoosh runs its synthetic shopper fleet entirely on Kernel. A customer picks the personas they want to test, hits run, and Kernel spins up a batch of browser instances that Squoosh’s agents drive through the site’s funnel, each one understanding the journey and walking through it. What used to take 30, 60, or 90 days of waiting for live traffic now comes back as a conversion report in under an hour.
“Kernel is the only solution we found that could actually scale with us. We can run thousands of synthetic shoppers in parallel, and we don’t manage any of that anymore. We just focus on our models.”
Tyler Dooskin, Founder, Squoosh
The team is now in the process of moving their agents to Kernel’s computer controls API so their synthetic shoppers run faster and more reliably as the newer computer-use models keep improving.
If you’re building agents that need to drive thousands of real browsers in parallel, with recordings your customers can actually trust, let us handle the infrastructure so you can focus on the model.
