How Rye scales agentic commerce with Kernel

Rye is the Stripe of agentic commerce. If you want to sell your own product online, you sign up for a Stripe account. But, if you want to sell someone else’s product online, you sign up for a Rye account and call their Universal Checkout API, which autonomously handles the end-to-end checkout process with any merchant on the internet. They’re trusted by Shopify, Visa, Amazon Business, and countless others. Plus, their co-founder just happens to be Justin Kan, the founder of Twitch.

Under the hood, Rye’s Checkout API uses AI-powered browser automations to get the job done. So it’s no surprise that they take the performance and reliability of their browser infrastructure quite seriously. 

Problem: Self-hosting infrastructure is hard

Prior to Kernel, Rye was self-hosting its browser infrastructure. At the time, there just weren’t many providers in the space to pick from, and they really needed performance. They wanted to run their agent on the same machine as the browser to cut down network latency and flakiness, which matters a lot for the overall execution time of their agent because you tend to need tons of CDP calls to do anything useful with a browser. So they leased a bunch of Mac minis and ran them in colocation data centers, hoping to get faster performance.

Of course, running your own infra like this definitely gets you control & performance, but it definitely doesn’t come free. You can’t just toss these boxes into an auto-scaling group, so the setup wasn’t very elastic and it forced them into a lot of upfront capacity planning. They also didn’t have nice APIs for provisioning; instead, they were maintaining a big repo of Ansible playbooks and configs for these machines, none of which actually helped them stand out from competitors.

On top of that, they were dealing with infra-related incidents every month. They had multiple outages tied to their in-house setup (running infrastructure is really, really hard). Each one meant a few hours of digging through logs, doing a postmortem, and shipping fixes, turning the whole thing into an ongoing time sink.​

“We wanted a fast end-to-end experience for our customers, but we really didn’t want to be in the browser infrastructure business. It’s really complicated to operate browser infrastructure. We have enough problems to solve.”

Sophia Willows, Head of Engineering, Rye

Ultimately, they decided it was time to buy, not build. And so they started searching for an alternative solution.

Solution: Selecting the best option in the business

Sophia’s team started the search by defining what “great” needed to look like: their agent had to run on the same machine as the browser to cut down on network latency and flakiness, new browser sessions had to come online quickly so cold starts did not affect their users’ experience, and the system had to scale elastically. To make the evaluation simple, they set a single north-star KPI: how fast their agent could run end to end, and how consistently it could do so without failures from network flakiness.

Sophia benchmarked Kernel against nearly every provider in the space, using the same acceptance tests they had relied on for their in-house setup. She focused on two metrics — end-to-end latency and reliability — and saw that other providers often had 4+ second cold starts, which is significant when an entire agent run takes only about 30 seconds. Kernel not only started faster but also delivered roughly 2x better end-to-end latency than the next-best provider in their internal benchmarks.​

Once the team was confident they had solved the core performance and reliability problems, they started looking at how Kernel could also help them ship a better, less frustrating checkout experience for their users. You see, one of the trickier challenges in agentic e-commerce is that the distance between a customer’s IP address and their shipping address is a strong fraud signal: if someone is buying from an IP in San Francisco but shipping to New York, that will typically raise their fraud score compared to a purchase from a New York-based IP shipping to New York. Kernel’s managed proxies, which support ZIP-level targeting rather than just state-level targeting like many other providers, suddenly became a powerful unlock for more precise fraud checks for their customers, letting them keep false positives down without relaxing their defenses.

With performance checked and fraud detection finally working with — not against — their agents, Rye started migrating their infrastructure to Kernel!

“Kernel is the fastest provider available, gives us agent identity features that are otherwise hard to build in-house, and removes our operational overhead. It was a no-brainer decision.”

Sophia Willows, Head of Engineering, Rye

Results: Absurdly fast agentic checkout

Today, Rye runs its AI-powered browser automations on Kernel, giving their customers a fast, friction-less checkout experience. At the same time, they have eliminated the hidden operational cost of having senior engineers babysit scaling infrastructure and unlocked a meaningful net reduction in overall infra spend, including $240K per year in direct engineering costs saved.​ All while getting the added benefit of our support.

“We love the Kernel team. They have been super responsive to our needs. I’m really happy with the level of support we get.”

Sophia Willows, Head of Engineering, Rye

If you’re currently building AI agents and dealing with browser infrastructure, let us help you too. We know a better way.