Harness^2: Recursive Agent Harnessing for an Open World

What it claims

Harness^2 is a plug-and-play framework that improves an LLM agent's execution harness (not the model) through recursive self-improvement (RSI), and is aimed at open-world tasks. It targets two problems the authors name for RSI. First, test-time tasks are novel, so no ground-truth verifier exists to guide harness edits. Second, RSI is usually task-specific and does not generalize to new scenarios.

How it works

Two connected levels of recursion:

• Task-level recursion. This level edits the harness for the task in front of the agent. A contrastive Propose-Probe-Compose step steers those edits in place of a verifier. The abstract does not explain what is proposed, probed or composed. The step can be scaled by running it in parallel or in sequence.
• Harnessing-level recursion. This level extracts experience from the task-level edits and uses it to update a persistent "continual improvement harness" that is carried to future tasks. This is the part meant to address generalization.

Evidence

• Coverage. Results span three professional benchmarks (WorkBuddy-Bench, JobBench, LAB) and four model-harness configurations built on GLM-5 and Gemini 3.5 Flash. The abstract does not say which harnesses were used.
• Baseline. Each model's own base harness, with a single Harness^2 step.
• Best-case gains. Up to +11.5 points on WorkBuddy-Bench, +4.5 points on JobBench and +11.8 percentage points on LAB. These are maxima across configurations, not averages.
• Scaling with budget. The single-step gains are reported as consistent across configurations, and scores rise further as the task-level recursion budget increases. This is the main point the catalog summary leaves out. Extra compute spent on task-level recursion buys more improvement, though the abstract gives no curve, no ceiling and no cost per step.

How to use it

The page offers the paper through the Hugging Face CLI:

hf papers read 2610.14224

To install the CLI if it is missing:

curl -LsSf https://hf.co/cli/install.sh | bash

No models or datasets are linked to the paper on the page, and the page does not mention code.

Limits

• Abstract only. The fetched page contains only the abstract. Per-configuration results, the definitions of the benchmarks and metrics, the probing mechanism, and compute or token cost per recursion step are all absent.
• Gains are best cases. All gains are "up to" figures, so the typical improvement is unknown.
• Baselines. Results are compared only with each model's base harness. The abstract reports no comparison against other RSI or harness-optimization methods.
• Generalization unquantified. The claim that the harnessing-level component helps future tasks is not backed by a separate number in the abstract.