The controller wrapping inference evolving-explainer

The shape of the AI product, not the model

The decision process external to the policy

A harness is the controller wrapping the transformer at inference: the loop deciding which tokens to feed in, which tools to call, what to remember, and how to route the output back into state.

If the model is the engine, the harness is the chassis. Two products can use the exact same engine and feel completely different — what the customer touches is the chassis.

The harness is the decision process that wraps a parameterized policy: it selects context, performs intermediate computation, and routes the policy's emissions back into state. The policy is one action among many.

Why scaling LLMs doesn't solve it

Why scaling models doesn't dissolve startups

Implications for self-improvement

Bigger models give you better tokens per call, not a better loop. The decision of when to call, what to feed, and what to do with the output lives in code outside the weights.

You can't bigger-model your way to a better product. The engine generates tokens; the chassis decides which tokens are even worth generating, and what surface they reach the user through.

Scaling the policy improves per-step behavior but does not constrain the harness's decision process. If the harness admits its own training signal, you can hill-climb the harness layer separately from the policy.

Meta-harnesses author Harness values from a typed grammar — the way an Architect emits UI manifests in meta-ide.

Meta-harnesses are the next move: products that shape themselves per user, per task. The product layout itself adapts.

The recursive case — meta-harnesses that author harnesses from a typed grammar — generalizes this to a search over harness families.

The shape of the AI product, not the model

If the model is the engine, the harness is the chassis. Two products can use the exact same engine and feel completely different — what the customer touches is the chassis.

Why scaling models doesn't dissolve startups

You can't bigger-model your way to a better product. The engine generates tokens; the chassis decides which tokens are worth generating, and what surface they reach the user through.

Meta-harnesses are the next move: products that shape themselves per user, per task. The product layout itself adapts.

The decision process external to the policy

The harness is the decision process that wraps a parameterized policy: it selects context, performs intermediate computation, and routes the policy's emissions back into state.

Implications for self-improvement

Scaling the policy improves per-step behavior but does not constrain the harness's decision process. If the harness admits its own training signal, you can hill-climb the harness layer separately from the policy.

The recursive case — meta-harnesses that author harnesses from a typed grammar — generalizes this to a search over harness families.