What's this about?
The paper's opening claim is the interesting one: agent capability is not determined by the model alone. The harness — memory management, planning strategy, action protocol, tool and skill orchestration — can dominate the model's contribution. Yet harness design is still manual, task-specific and unscalable. JIT-Agent evolves the harness at runtime instead. 47 upvotes on HuggingFace.
What is a "harness"?
Everything wrapped around the model that turns it from a text predictor into something that gets work done: the loop that calls it repeatedly, the tools it can invoke, what it remembers between steps, how it plans, what it does when something fails. Two teams using the identical model with different harnesses get very different results — which is why "which model is best" is often the wrong question.
Why does this matter to you?
- It is the formal version of an argument that ran through HN this week in the "Harness Engineering" thread, and through the small-models discussion.
- If the harness dominates, then evaluating models in isolation systematically mismeasures them.
- Practically: it suggests investing in your scaffolding before upgrading your model.