Pax8 · AI agent platform · 2026
I designed the console IT partners use to deploy and run AI agents across their clients, then built it as working code with Claude. Some of that code went through engineering review into production.
Static designs showed what the console looked like, not how it behaved.
Design reached engineering as Figma files and walkthroughs, and engineers translated them into working interfaces. Interactions, responsive behavior, and multi-step workflows didn't come through fully, so there was back-and-forth during implementation, and some decisions had to be revisited once building started.
It’s a complex console with interconnected workflows. Working in code meant I could make decisions while using the actual interface, find problems earlier, and iterate without recreating every state in Figma.
Design used to end at a Figma handoff. Here it ended in working code that engineering could review.
Some of my code became the starting point for production.
Engineers reviewed what I pushed and brought parts of it into the production codebase, including work on the deployment and agent experiences, which the team later merged into one workflow.
Because I was working in code, I took on the detailed design refinement myself: layout, interactions, visual consistency, the small things that usually take rounds of back-and-forth. That left engineering to focus on integration and production readiness.
Speed created a new problem: several working prototypes, and no clear way to tell which one was the current direction.
Different people were exploring different versions in parallel. It got hard to tell which decisions had been made, which ideas were still exploratory, and what we were actually building.
With the team, I set up one working prototype as the source of truth for how the product behaved, and a Miro board of screenshots as the source of truth for decisions, because not everyone reviewing the work reads code. PMs, content designers, and designers commented there and divided up ownership.
The goal wasn't to slow experimentation down. It was to make fast experimentation possible to coordinate, so nobody duplicated work or built on conflicting assumptions.
I took it on myself to share the workflow with other designers. Nobody had to use it, so I made it easy to pick up.
I paired with other designers, ran demos, and wrote a guide in Confluence showing how I used Claude to build and refine working prototypes. Every designer had their own way of experimenting with AI, and adopting any of it was their call. A few designers started using parts of the workflow in their own work.
The bigger shift was in how people saw code: as another way to explore and communicate a design, not only something that happens after design is done.
AI made producing solutions faster. It also made alignment and a shared source of truth more important.
When anyone can build a working prototype in an afternoon, the hard part moves from making things to agreeing on which thing we’re making.
The full case study is password-protected until launch.
The product is unreleased, so the screens, research, and design system live in a private case study. It covers how the brief changed, designing for an agent that works without a screen, keeping code and Figma in step, and checking the Beta against the real build.
Email me for the password, or for a live walkthrough.