What's this about?
This paper (Jul 14) argues a modern agent's capability depends not only on its model but on its harness — the code that constructs prompts, manages state, invokes tools, coordinates execution — and proposes methods for making these fast-evolving harnesses readable, navigable, and editable.
What exactly is the harness/model split?
The model is the neural network; the harness is everything wrapped around it. Two products using the *same* model can behave wildly differently because of harness quality — which is why Claude Code, Codex CLI, and Grok Build compete on harness engineering as much as model quality. Harnesses have become sprawling codebases (Grok Build: 844k lines) that change weekly; treating them as first-class engineering artifacts is overdue.
Why is it worth noting?
It's the academic echo of today's news cycle: Grok Build open-sourcing its harness, the "harness that can do anything" HN debate, and Logan Kilpatrick's recent "the model eats the harness" thesis — all the same argument from different sides.