I thought I was working on the cutting edge. Not until our product was drivable by an agent.

I thought I was working on the cutting edge. I wasn't. Not until our product was drivable by an agent.

We built an MCP server. Looked like an integration task.

Exposing a good product over MCP isn't hard conceptually. It's also not a weekend. We spent weeks tuning, and a couple of months before an agent could get through a real task without hand-holding. Descriptions that say what happens and what fails. Operations that do one thing. Results an agent can actually act on instead of a success flag.

Two things came back, and neither was distribution.

First, watching what people did with it. Our own team started driving the product from their agents and built stuff I'd never have specced. Customers who live in agents did the same. You find out what your product is for by watching what people build out of it. And you can't run that experiment until the pieces are callable.

Second one I didn't see coming. Once every piece is drivable on its own, the question changes. Not what do we build next. Given these pieces, what's possible now that wasn't?

I don't think you can ask that honestly beforehand. What you imagine your product could be is stuck inside the interface you already have. Mine was.

The tuning is the real cost. Standing the endpoints up was quick. Getting an agent through a full task without babysitting took a lot longer.

If you've done this, curious whether you got the same shift, or whether I'm overrating it because a bunch of other things happened at the same time.