how anything builds agents that live in the cloud with KERNEL

anything has raised $11 million at a $100 million valuation with a mission to let anyone build things that people will love. their latest product is skydive, which allows agents to do anything a person can do in their browser.

for this experience to work as intended, users needed to be able to see what the agent was doing in real time and interact with the browser while the agent took action. after building a dedicated testing playground to benchmark against four other browser providers, marcus and the team chose KERNEL browsers to power skydive’s live view experience, streaming at 24 frames per second (fps), which builds trust with the user and makes the experience feel natural.

“most people’s work happens inside the browser. we wanted skydive’s agents to work there, too.”

marcus lowe, anything, co-founder & ceo

problem: latency was blocking skydive’s live view experience

a key part of the skydive user experience is live view, a feature that lets users watch an agent work in a browser and take over at any point. for example, if a skydive agent is triaging support tickets, a user can watch it trace a bug through the logs in real time, then step in to review a reply before it goes out.

while live view consists of a sandbox running an agent and a browser running somewhere else, it’s wired together in a way meant to feel like one seamless machine. this seamlessness is critical; if it breaks, users stop trusting the agent.

skydive needed more than anything’s previous browser infrastructure setup, hosted on browserbase, could provide. the team believed the best solution was to build something custom, so they tried home-rolling their own infrastructure instead of partnering with a provider. however, when anything built the initial version of skydive’s browser experience, live view was too slow. the team traced the lag to the infrastructure. each agent’s compute ran inside a sandbox (an isolated environment where agents can die, restart, or get replaced), but it had to reach separate infrastructure on a different network to control the browser. they believed that distance caused the lag.

their first attempt to fix the lag was to remove the distance entirely by running chromium inside the same sandbox as the agent. it didn’t work. the team suspected the network layer wasn’t built for that kind of load.

“we ran a small media server inside the sandbox that also piped out the video feed. it just sucked.”

marcus lowe, anything, co-founder & ceo

the next idea was to deploy the sandboxes across multiple regions, so the browser could sit closer to users no matter where they were. but skydive’s sandbox provider only runs servers in the us west, europe, and asia regions. the team considered switching to a provider that offered more regions, but it also meant porting over large amounts of infrastructure the team had already built for uncertain improvements in latency. the tradeoff didn’t seem worth it.

solution: skydive built a shared, fast browser experience, with KERNEL’s infrastructure

instead of trying to build the infrastructure themselves, the team building skydive decided to find another third-party provider for browser infrastructure. they built a playground to compare providers directly by pointing each one at google, linkedIn, and github, then layering on ip and stealth-detection tests. skydive benchmarked every browser infrastructure provider, and KERNEL won on speed and interaction.

“even without knowing the numbers going in, we already felt the difference. it looked like KERNEL was streaming at 24 fps, and everyone else was streaming three. that feeling of using KERNEL is what stood out.”

marcus lowe, anything, co-founder & ceo

while KERNEL’s browsers run in us-east, farther from skydive’s san francisco users than the setup they’d already tried and abandoned, it still felt faster.

working with KERNEL helped the team discover another benefit it hadn’t anticipated. skydive’s sandboxes cycle constantly, restarting or getting replaced, and each cycle risks losing whatever state the browser had built up. KERNEL’s profiles feature helps preserve the same browser profile, so it follows the agent across every session, regardless of which sandbox is running it.

results: skydive shipped cloud browsers and avoided months of engineering work

“without KERNEL, we couldn’t have shipped skydive’s live view at all. KERNEL saved us from needing to port our infrastructure to a different provider, which would have taken somewhere between a week and three months of engineering work.”

marcus lowe, anything, co-founder & ceo

with this implementation, anything is able to create a snappy, responsive user experience with skydive. users can watch an agent work inside a browser (i.e., one that starts quickly and holds up under interaction), and the infrastructure connecting their sandbox to KERNEL’s browsers stays invisible.

if your agents need to earn a user’s trust in real time, browser infrastructure is what makes or breaks it. we can help.