I Built a Model Router I Didn't Need

I finally tried Pi.

First, I spent the Fourth of July building something else.

My workflow ran through Claude Code. The vault was the brain, skills covered frequent tasks, and the Command Center told me where to pick up work.

I was terrified of rebuilding this for another harness.

Pi wasn’t new to me. What changed was that I finally understood what these harnesses were capable of. From a distance, Pi looked like flexibility and freedom.

I also wanted to try other models. Model routing was all over X, and it came up routinely in developer meetings at work.

But I treated trying Pi like a migration project before I’d even tried it.

So I started Baton. I wanted to move seamlessly between harnesses, untethered from any one ecosystem.

I built a runtime that routed work to swappable model workers. The handoff worked.

But by the time I looked up, Baton was eating the hours I’d set aside for Buildtold, the product sitting at the top of my Command Center.

I tried Pi anyway.

I told the agent which workflows I needed and asked it to look at how I already handle them. The ask was roughly:

I use [YOUR HARNESS] and I don't want to rebuild my setup here.
Look at how my workflow is configured there:
- instructions
- skills
- project conventions

Report back on what can be pointed at directly and what needs an
equivalent. Don't change anything yet.

It came back with a plan that pointed Pi at my Claude directories. Nothing moved. The vault and my instructions transferred or had equivalents waiting.

I still have kinks to work out around permissions and modes, but I didn’t need to rebuild my flow to get started.

The model was never the dependency, and neither was the harness. What I assumed would hurt was moving my skills and configuration between tools. It didn’t hurt. It was easy.

Baton solved a problem I didn’t have.

Every setup I scroll past feeds my agentic FOMO. Another model, another harness, another layer on top of the last one.

You don’t have to flip every switch the day you pick up a new tool. I still haven’t touched half of what Herdr and Pi can do.

Some of it helps, and some of it solves problems I don’t have.

Every step has a cost. Before long, I’ve got fifty steps between an idea and ten lines of code.

I started learning guitar late, and my son started young. He picks it up and plays it wrong at full volume. I sit there trying to get it right before I make a sound.

The tools were never holding me back. I’m still tuning.