Nava Labs is the company’s philanthropically-funded division that focuses on prototyping policy and system changes in government programs and advocating for their adoption. Since 2015, federal, state, and local government services have trusted Nava to solve highly scrutinized technology challenges.
Last June—as part of a project to empower caseworkers with AI support tools—Nava Labs received a $1.5 million grant from Google to partner with California-based community partners on developing and testing a suite of generative AI agents designed to autonomously navigate multiple databases and benefit portals to fill out applications, which caseworkers can then review, revise, and submit.
Problem: screens getting in the way of trust
Addressing this problem required working closely with county caseworkers, who often meet with families walking into a county office simply saying “I need help.” These families usually don’t know which benefits they qualify for or what information they’ll need, and caseworkers help by guiding them through long, high-stakes forms—many of which ask for deeply personal details.
Between them sits a large desktop screen that can make the interaction feel transactional and reduce trust, especially when the person in need is already hesitant to share personal information.
Nava Labs wanted to answer a simple question: can we remove the screen as a barrier and automate the handling of paperwork in the background so caseworkers can focus on the human in front of them? That meant understanding how to navigate these online applications, keep up with the conversation, and reliably fill out forms across many different government systems.
At the same time, they faced a second constraint: traditional approaches like RPA and static Playwright scripts weren’t a good fit. Government forms change frequently on the client side—links move, layouts update—making deterministic scripts brittle and expensive to maintain.
“We wanted something robust that could work across benefit applications, without constantly rewriting Playwright scripts every time a form changed.”
Foad Green, Senior Software Engineer, Nava Labs
Solution: Selecting Kernel over self-hosting
When Nava Labs started designing their system, browser agents were still relatively new, but they were interested in exploring it. Browser agents could give them the robustness they were looking for by autonomously adapting to any UI changes.
Their first version worked, but it didn’t scale well. They also didn’t want to be in the business of constantly growing and managing a fleet of browsers and dealing with performance bottlenecks. On top of that, their “Live View” or takeover experience—where a caseworker can manually review and adjust an application inside an embedded browser—needed to feel natural and responsive. Their homegrown version had gaps: special characters sometimes broke interactions, mobile support was incomplete, and mouse and keyboard input felt noticeably laggy.
“We wanted to focus on the product and user experience, not on running a fleet of browsers.”
Foad Green, Senior Software Engineer, Nava Labs
So instead they turned their heads towards managed solutions. Nava Labs evaluated several browser providers and compared them against their self-hosted implementation. Kernel stood out on three fronts: scalability, real-time live view performance, and built-in capabilities for authentication.
Results: 55% infra cost reduction, 3x faster build times
Today, Nava Labs runs its AI agents on Kernel, giving caseworkers and families a far more human experience. At the same time, switching to Kernel eliminated the need to manage 3 Docker containers per deployment. This change not only tripled their build speed—from 12 minutes to just 4—but also reduced overall GCP infrastructure costs by 55%.
With Kernel’s Live View, Nava Labs closed the remaining gaps. Special character handling and page interactions now work as expected and caseworkers can scroll, click, and type as smoothly as if the page were running locally.
“The development experience with Kernel was noticeably better than with other providers, and the support has been incredible.”
Foad Green, Senior Software Engineer, Nava Labs
If you’re building AI agents and wrangling browser infrastructure, we can help. There’s a better way.
