Harness is Gradually Absorbed by the Model

The “bitter lesson” applied to the age of agents: a general method will eventually win out, but every stretch of the road to that “eventually” is paved with Harness.

Example: interruption and insertion of utterances—behavior previously orchestrated by external Harness—is now built into the interaction model.

Three reasons why Harness won’t disappear:

  1. training takes months (the model will wait, business won’t);
  2. the model doesn’t internalize all constraints of a specific business;
  3. each new generation of model opens a new frontier of capabilities where the model is least stable.

Harness migrates with the model to new frontiers.

Related: Harness-engineering, Model-agent coevolution