Researchers publishing on Hugging Face have introduced Harness², a framework aimed at improving how large language model agents operate in open-world settings. The work argues that an agent's capabilities depend not only on its underlying model but also on its execution harness — the surrounding scaffolding that governs how the model acts.
Recursive self-improvement (RSI) has been proposed as a way to automatically evolve such harnesses, but the paper identifies two obstacles to deploying it in the open world. First, open-world tasks are typically unseen during harness design and novel at test time, so there are no ground-truth verifiers to reliably guide online edits. Second, RSI processes tend to be highly task-specific and fail to generalize to new scenarios.
Harness² addresses these issues through two connected levels of recursion. Task-level recursion refines the current task's harness using a contrastive Propose-Probe-Compose step that guides edits without ground-truth verifiers, and can be scaled further through parallel or sequential recursion. Harnessing-level recursion then extracts experience from those edits to update a continual improvement harness, helping the agent adapt to future tasks.
Across three professional benchmarks and four model-harness configurations, the authors report that a single step consistently improves GLM-5 and Gemini 3.5 Flash over their base harnesses. Gains reached up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB, with further improvements as the task-level recursion budget increases.