A harness is the shell around a model — the prompts, tools, memory, and UI that decide what the model sees and what it can do. Most teams treat the harness as fixed scaffolding.
The interesting move is to author the harness itself: an LLM emits a typed UI manifest per phase, and the editor reshapes around the task at hand.
Planning looks like a research scratchpad. Drafting collapses to one editor. Review becomes a diff. Publishing produces a reader-harness that adapts to each reader's prior.
A harness wraps a model with prompts, tools, memory, and UI. Teams treat it as fixed scaffolding — it shouldn't be.
An LLM (the Architect) emits one Manifest per phase. The Renderer walks it. Pin/dismiss feedback survives refines.
If the model's harness is the policy's environment, then learning a harness is meta-learning: the outer loop optimizes the shell that constrains the inner policy.
Trace events power offline eval. Pinned panels become a per-author preference summary that biases future Architect output.