HarnessPro

What it is

HarnessPro is a method reportedly published by researchers from Tsinghua University and Zhipu, with the paper posted to arXiv in October 2026. A harness is the outer framework that wraps a model and makes it perform tasks, including prompts, tools and execution procedures. Recursive self-improvement approaches, in which agents improve their own harness, existed before, but the paper points out that such approaches often raise execution cost along with performance. The core of HarnessPro is revising the harness while weighing performance and cost together.

How it works

First, the agent runs tasks with the harness and records the execution process (Trajectory). Agents assigned specialist roles diagnose these records, draw up improvement plans with different priorities such as performance-first or cost-first, and create new candidate harnesses. Candidates are evaluated with LLM-generated tests rather than ground-truth answers or official tests; this is called a label-free approach because no ground-truth labels are needed. Then three harnesses are selected — cost-focused (lite), balanced (medium) and performance-focused (heavy) — and passed to the next iteration, continuing a loop of revising based on execution results.

Why it matters

Because it improves agents by revising only the harness without retraining model weights, it is a relatively lightweight direction to try. Rather than a single 'best' harness, it offers three for different budgets, so cost-sensitive services and performance-sensitive services can each choose. The paper states that on Terminal-Bench 2, SWE-Bench Pro and ProgramBench, using GPT-5.5 with Codex, heavy mode raised the pass rate by an average of 2.0 percentage points over the comparison method (RHO) and cut cost per task by an average of 26.1%. It said these advantages also held for GLM-5.2 used with Claude Code.

Things to know

As of `today`, HarnessPro is a research paper under review, not a finished commercial product. The performance and cost figures are results reported by the researchers themselves and are based on specific benchmarks and comparison methods, so whether the same effect holds in real work environments needs to be checked separately. Because evaluation uses LLM-generated tests, there is also a risk that the harness could be 'improved' in the wrong direction if those tests are weak. Since the agent changes its own execution environment, it is safer in practice to also have a mechanism for humans to review the changed harness.

Example

This site's news item 'Tsinghua University and Zhipu's HarnessPro: agents revise their harness by reviewing their own execution records… "26.1% lower cost per task" (reported)' covered this term. For example, a team running a coding agent could picture using this approach to split the work: a cost-saving lite harness for large-scale internal automation tasks, and a performance-focused heavy harness for difficult bug fixes.

Where it came up on this site

• briefing Tsinghua University and Zhipu's 'HarnessPro': agents revise their harness by reviewing their own execution records… "26.1% lower cost per task" (reported) 2026-10-31

Related terms

Harness AI agent Coding agent Recursive self-improvement Benchmark LLM-as-a-Judge

References

• Tsinghua and Zhipu's HarnessPro Lets Agents Rewrite Their Own Harness, Cutting Task Costs 26% (Pandaily)
• How China's AI Labs Are Chasing Recursive Self-Improvement Without Touching Model Weights (Pandaily)
• HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs (arxiv.org)
• Tsinghua and Zhipu's HarnessPro Lets Agents Rewrite Their Own Harness, Cutting Task Costs 26% (Pandaily)