Harness²
What it is
Harness² (from Google Research and Northwestern) is a framework for improving a frozen, non-retrained agent by rewriting the harness around it instead of fine-tuning the model's weights. It runs two nested recursion loops: a task-level loop that proposes, probes, and composes candidate edits to the harness for the task at hand, and a harnessing-level loop that distills lessons from those episodes into a continually-improving harness carried forward into future tasks.
Why it's emerging now
Once a harness is treated as a designable artifact in its own right (see "Harness engineering" below), the natural next question is whether an agent can improve its own harness automatically — and keep the improvement across tasks rather than relearning it each time. Google's reported gains (+11.5 points on WorkBuddy-Bench, +11.8 on LAB) with zero weight updates make a strong efficiency case for that approach. A separate, differently-authored paper on the same idea — Recursive Harness Self-Improvement, from Sakana AI and UC Berkeley — landed in roughly the same window, suggesting multiple groups converged on recursive harness self-improvement independently rather than one team coining a one-off term.
How it relates to what you know
It builds directly on two entries already on your ledger: "Harness engineering" treats the harness as a first-class designable artifact, and "Meta-harness" is about orchestrating across multiple harnesses. Harness² is narrower and more mechanical than either — a specific two-level recursive algorithm for making one harness self-improving across tasks, closer to a training procedure for the harness than to a design discipline or a cross-harness orchestration layer.