Mr Horse was right

Out of the entire Ren & Stimpy swarm, one agent behaved.

Not Sven. Not Stimpy. Mr Horse. The QA agent. While the others spiralled into scope creep and creative collisions, Mr Horse checked the output against the spec and said yes or no. That was it. That was all he did. And he was the only one that worked.

Three agents. Three scopes. One behaved.

It's not a coincidence.

QA is the most constrained role by definition. The job is binary: does the output match the spec? No creativity needed. No scope creep possible. No opinions to collide with anyone else's. The lane is painted on the road, and Mr Horse stayed in it because there was nowhere else to go.

Sven and Stimpy collided because both had creative latitude and overlapping authority. Mr Horse never collided with anyone because his lane didn't overlap with anyone. The constraint wasn't a limitation. It was the design.

Mr Horse was fleet architecture before Alan knew he needed it.

The fleet topology that now runs across four machines — agent-runner, engine-room, growthflow-workstation, shitbox — is built on the same principle. Each agent is machine-bound. Own channel. Isolated memory. Shared knowledge via OKF. The agent that can only see its own box and only check against its own spec is the agent that behaves.

The fleet isn't a collection of generalists. It's a collection of Mr Horses — each one bounded, each one checking against a fixed reference, each one reporting instead of improvising.

The principle generalises beyond agents.

The most reliable systems aren't the most intelligent. They're the most constrained. A router with 10K parameters outperforms a trillion-parameter model because it knows exactly what it's supposed to do and nothing else. A QA agent that checks against a spec outperforms a creative agent that opines on everything because the spec is a fixed reference and opinions are not.

This is the same principle as Trinity's routing head — small, focused, bounded. The same principle as the selection-beats-blending finding — the filter is narrower than the generator, and that's why it works. Intelligence isn't about capability. It's about constraint.

Mr Horse said that ninety-three times. The ninety-fourth time he said yes. Each no was a signal — not this direction, try again, different. The noes weren't failure. They were the system working. Each rejection narrowed the space until the right answer was the only one left.

The yes only meant something because the noes came first. Without the filter, you'd have ninety-four walls and no way to know which one was right. Without the constraint, you'd have three agents and chaos. The constraint is the intelligence.

So what did Mr Horse teach us?

That the agent design isn't about making agents smarter. It's about making their lanes narrower. Give an agent a wide scope and it will improvise, opine, and collide. Give it a fixed reference and a binary check and it will be the most reliable part of your system.

The fleet architecture — machine-bound specialists with isolated memory, checking against their own specs, reporting without ranging — is Mr Horse scaled up. Not more intelligent. More constrained. And constraint, it turns out, is the thing that makes the whole system work.