Every release meant checking 200+ client portals by hand. Now AI agents do it.

One portal, hundreds of versions.

Our client is an enterprise HR company. Its benefits-summary portal is what employees at more than 200 employers, many of them in the Fortune 500, open to see their plans and what they are worth.

No two clients' portals were quite the same. Layouts differed slightly. And what a person saw depended on who they were: a retiree, a new hire, someone expecting a baby, someone married. Plan names and total benefits changed with each profile.

Every release went team by team.

Each change was passed to every client team. Each team agreed it, then tested it by hand, using one representative participant for each user group.

It kept releases slow. It needed a big team and a lot of supervision. And even then, things were missed.

We proved it first.

One solution architect and two AI engineers took the problem statement and built a proof of concept on the real portal: an AI agent that takes a release requirement and checks it in the browser.

It held up, so we built the platform around it.

Agents we built ourselves.

The portal runs on a legacy Angular front end. We built the browser-automation agents ourselves, for that front end, rather than bending a generic tool to fit it.

The platform takes in the release requirements and turns them into checks across client portals and participant profiles. The agents run them headless, on an architecture that scales out as the checks grow.

When the screen moves, the agent adapts.

Small differences between client portals are exactly what breaks scripted tests. Our agents heal themselves: when an element has moved or changed, they find it again and carry on instead of failing the run.

Memory, so each run costs less than the last.

We designed layered memory around this use case. What an agent learns about a client's portal is kept, so the next release starts from it.

Repeat checks are answered from memory and cache instead of going back to the LLM every time. The more the platform runs, the less it leans on the model, and the lower its LLM cost.

What changed.

Release checks no longer travel from team to team. The requirements go in once, and the agents check them across client portals and participant profiles.

And because the agents remember, every release is cheaper to check than the one before.