The performance of LLM agents hinges not only on their underlying models but also on the execution harnesses that orchestrate them. Recursive self‑improvement (RSI) promises automatic evolution of these harnesses, yet deploying RSI in open‑world settings encounters two core obstacles: first, tasks are usually unseen during harness design and novel at test time, leaving no ground‑truth verifiers to reliably steer online edits; second, existing RSI pipelines are tightly coupled to specific tasks and fail to generalize.\
Harness^2 addresses these issues with a plug‑and‑play, two‑level recursion framework. At the task level, a contrastive Propose‑Probe‑Compose cycle refines the current task's harness: candidate edits are proposed, their impact is probed by measuring performance differences on sub‑tasks, and the most effective edit is composed back into the harness. This process operates without ground‑truth labels and can be scaled via parallel or sequential recursion to deepen the edit search.\
The harness‑level recursion then aggregates experience from all task‑level edits, continuously updating a continual‑improvement harness that equips the agent to adapt more efficiently to future tasks. Across three professional benchmarks—WorkBuddy‑Bench, JobBench, and LAB—and four model‑harness configurations, a single recursion step consistently outperforms the base harnesses for both GLM‑5 and Gemini 3.5 Flash, yielding gains of up to 11.5 points on WorkBuddy‑Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB. Further improvements accrue as the task‑level recursion budget increases.\
Review: Harness^2 demonstrates that a contrastive edit mechanism combined with experience accumulation can effectively optimize harnesses in open‑world scenarios lacking ground‑truth verification, highlighting the practical viability and scalability of recursive self‑improvement for LLM agents.