What's this about?
A benchmark that measures whether an LLM can improve the "harness" around it — the prompts, tools, control flow, and memory — rather than only its raw weights.
What exactly is a "harness"?
The harness is everything wrapped around the model that turns it into a working agent: the system prompt, the tools it can call, how its steps are sequenced, and how it stores memory. A growing thesis says the harness matters as much as the model; this paper makes "can the model tune its own harness?" a measurable task.
Why it matters
If models can optimize their own scaffolding, agent systems could self-improve without retraining — and we now have a way to score that ability.