Vantara Agent Studio

The problem

The Vercel PM (Agent Platform) role asked for someone who understands where enterprise agent UX is heading. I didn't want to write another cover letter about it. I wanted to build something instead — a real product that shows the PM thinking, not a doc that describes it.

My hypothesis

If I can design a credible enterprise agent-building workflow in a day — with real AI responses, a dashboard, and version history — I can show the product judgment that matters for the role without a slide deck.

What I built

A 4-step wizard that takes an enterprise use case to a deployed agent: collect context, generate Claude-powered clarifying questions, preview agent output, simulate deploy. Built on Vercel AI SDK edge runtime with Upstash Redis for persistence. The scenario: a healthcare org managing 500 employees across 15 practices, with three agent types — Leadership (monthly updates), Regulatory (federal policy monitoring), Operations (patient volume). Demo mode lets a hiring manager walk through the full wizard without any API setup. Try it: vantara-agent-studio.vercel.app/build?usecase=comms&demo=true

What broke

Demo mode revealed something: the experience with real Claude responses and with pre-baked ones felt almost identical — which means the UX is doing real work independent of the model. The enterprise vertical also taught me that agent use case framing is everything. The same capability reads completely differently depending on the job it's being asked to do.

What I learned

Structuring the wizard as a 'prompt for the human' — not a configuration UI — was the right call. Non-technical users don't know how to specify what they want from an agent. The clarifying question step isn't about collecting data; it's about helping the user articulate the problem.

If I kept going

This stays a demo unless I get the job. If I kept going: A/B test the clarifying question step — fewer, better questions probably outperform the current model. Version history with semantic diffs (what actually changed about the agent's behavior, not just a timestamp) would be the real product bet.

What was built

Prototype

Landing: build agents your team can use — no engineering required

The 4-step wizard end-to-end: starting point → clarifying questions → preview → deployed

Dashboard: 3 agents deployed, version history, schedule, and status