CASE STUDY · AGENTR · PIVOT TO AN AGENTIC AI PLATFORM
The two-week read that changed what we were building
AgentR (Ruzo) · Founders' Office · Product & Growth · April – November 2025
“Two weeks in, I saw we were building plumbing for someone else's product. I made the case to build the product itself — and then designed and shipped its v1.”
A clever piece of infrastructure — and a thin place to stand.
When I joined AgentR, the product was a clever piece of infrastructure based on the trend of the market. It was a web app that let you assemble your own MCP server. You'd switch on the integrations you wanted (Gmail, Notion, whatever), and the app would expose them as a single server you could plug into an MCP client like Claude Desktop or Goose. The team's whole motion was building more integrations: 20 were live when I arrived, and that climbed past 75.
In plain terms: we were the wiring that connected your apps to someone else's chat interface. The product was in beta — pre-revenue, around ten users, all connections of the founders or investors — and the company was raising.
That was the picture I walked into. It took about two weeks to realise it wasn't enough.
Stop building plumbing. Build the destination.
Here's what I saw. An MCP-server aggregator is a thin place to stand. You're exposing tools into a client you don't own — Claude, Goose, etc. — which means the experience, the surface the user actually lives in, belongs to someone else. And "more integrations" is a treadmill: once the top apps are connected, each new one matters less, and nothing about the position is defensible. The value in this category wasn't going to accrue to whoever had the most integrations. It was going to accrue to whoever owned the client experience — a single place to work, with the best models from every top provider, where the integrations live inside the product instead of being piped out to a competitor's window.
So my read was: stop building plumbing, build the destination. A full agentic platform — our own client, our own ecosystem — was the product worth owning.
The founders didn't agree at first. For about two weeks they were hesitant, and that's fair — I was arguing to build a parallel product on a team of five mid-raise. What turned it was the treadmill making my own case for me: once all the top apps were integrated, "add more" visibly stopped paying off, and a wrapper around the experience became the obvious next move. I made one more argument that mattered — that if we were going to build it, we should build it in a way that kept our options open. This wasn't a de-scoping of the existing product. The MCP-server web app stayed live in parallel for the next six months. Ruzo was a new bet set up alongside it that grew into the main thing — which meant we redirected our future without throwing away what already worked.
That was the call. The thesis was mine; the founders' role was to be convinced, and then to approve.
Ruzo — a full agentic platform, designed and shipped as v1.
I owned the product end of this: the PRDs, the wireframes, the UX flows, the feature prioritisation, and the QA. The founders approved and the engineers built from my stories. The shape of v1 was mine.
What I owned
What a user could actually do at launch: toggle apps on, and within each app toggle the specific tools they wanted. Each enabled tool got exposed through a server that ran locally — set up on your machine and run inside clients like Claude Desktop and Goose. It was the same underlying power as the old product, but now wrapped in an experience we controlled and could push in any direction.
The architecture direction was mine too, by insight if not by hand. I pushed for moving from MCP integrations to a LangGraph-based Agentic system with direct integrations. The reason was concrete: MCP capped how many tools could be meaningfully called in a single response, and that ceiling limited how deep an agent could go. LangGraph broke that open — it gave us access to genuinely different agent types like ReAct and CodeAct, and far more depth in multi-step work. I identified the limitation and called the direction; the engineers confirmed it was right and built it.
The part I'm proudest of is the standard. v1 had to do a lot of things right just to earn a look. Our users already had ChatGPT and Claude Desktop open — to get them to switch, the product couldn't have rough edges. It had to be well laid out, robust, and complete from the first day. That bar — usable and marketable from day one, not after three patch cycles — was the one I drove into the build. Getting the small details right under that constraint is the work I rate most in this whole project.
I'll be straight about where this lands.
I'll be straight about where this lands, because the value here is in the judgement and the craft, not in a metric I didn't stay for.
I moved on from AgentR before go-to-market ramped up. The funding round was already in, and the full GTM push hadn't begun on all cylinders when I left — so I can't point you to adoption or revenue numbers for Ruzo, and I won't invent them.
What I can stand behind: I read a weak strategic position in two weeks, made the case to change it against initial founder resistance, and then designed and shipped a complete, polished v1 of the platform that replaced it — owning the product work from PRD to QA, and driving the architecture call that gave it real agentic depth. The pivot was the right call on the merits. The build was mine to lead, and I led it to a standard I'd defend in any room.
“The most valuable thing a PM can do early isn't ship faster — it's notice, before anyone else does, that the whole team is building the wrong thing for the right reasons. Then make the case, and build the right thing well enough that switching to it feels obvious.”