A harness is the controller that wraps the transformer at inference time: the loop that decides which tokens to feed in, which tools to call, what to remember, and how to respond to outputs.
The transformer is policy. The harness is everything else. Scaling the policy doesn't fix harness shape — that's a different optimization problem entirely.
If the model is the engine, the harness is the chassis. Two startups can use the same engine and ship completely different products — the chassis is what the user feels.
This is why scaling models doesn't dissolve startups. The harness — the product surface, the memory, the way the agent loops — is where the moat lives.
Define the harness H as the controller wrapping a parameterized policy πθ. H selects context, performs intermediate computation, and routes the policy's emissions back into state. The harness is, formally, an MDP whose action space includes invocations of πθ.
Scaling θ improves the policy's per-step behavior but does not constrain H's decision process — they are orthogonal optimization targets, and recent results (RLM hill-climbing) suggest the harness layer admits its own training signal.