AnswersPublished15 sources
HarnessPro: How It Optimizes AI Agent Performance and Cost
HarnessPro iteratively changes an AI agent’s instructions, tools and runtime policies, then offers Lite, Medium and Heavy versions for different cost–performance trade offs. Rather than optimize for one universal configuration, HarnessPro uses synthesized tests and Pareto based selection to retain options for lower...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- HarnessPro iteratively changes an AI agent’s instructions, tools and runtime policies, then offers Lite, Medium and Heavy versions for different cost–performance trade offs.
- Rather than optimize for one universal configuration, HarnessPro uses synthesized tests and Pareto based selection to retain options for lower cost, balance and higher performance.[3][11]
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is HarnessPro, released by researchers from Tsinghua University, Zhipu, Shanghai Jiao Tong University, the University of Warwick and Ku. Article summary: HarnessPro is a self-supervised framework for recursively improving an AI agent’s **harness**—the instructions, tools and runtime policies around it—while optimizing both task success and execution cost.[2][16] Its disti. Topic tags: general, academic, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake num
HarnessPro is a label-free framework for recursively improving an AI agent’s harness: the instructions, tools and runtime policies that shape how it works. It evaluates candidate changes with synthesized tests and retains configurations across different performance–cost trade-offs, rather than treating one setup as best for every use case.3
11
What an agent harness does
A harness is the surrounding setup that guides an agent’s work—not just the underlying model. It can include instructions, available tools and runtime policies. HarnessPro’s premise is that changing this setup can affect both task performance and execution cost.11
How HarnessPro improves a harness
HarnessPro uses execution traces to identify patterns in failures and costs, then generates candidate changes to the harness. Those candidates are evaluated with synthesized tests instead of relying on ground-truth solutions or official tests during optimization, according to the project README.11
The process is recursive: candidate harnesses can be evaluated and refined in further rounds. That makes HarnessPro a method for evolving the agent’s operating setup, rather than a change to the model’s weights.3
11
Why it keeps Lite, Medium and Heavy options
A cheaper harness may be preferable when cost matters most; a more capable one may be worth choosing when performance is the priority. HarnessPro’s Pareto-based selection preserves three operating points: Lite for lower cost, Medium for balance and Heavy for higher performance.11
The practical idea is to make the trade-off explicit. Teams can select among the available options according to their deployment needs instead of assuming that the configuration with the highest score is always the right choice.11
What the reported benchmarks show
A project results page reports that, with GPT-5.5 and Codex on Terminal-Bench 2, HarnessPro Heavy achieved an 80.8% pass rate at $0.72 per task. The same comparison lists RHO at 76.3% and $1.09 per task.16 Separately, research summaries report an average pass-rate gain of 2.0 percentage points and a 26.1% reduction in per-task cost versus RHO for the reported evaluation setup.
7
20
These figures describe particular model, harness and benchmark combinations. They should not be read as evidence that HarnessPro will produce the same gains on every task or deployment. The project presents multiple variants precisely because performance and cost can pull in different directions.11
16
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "HarnessPro: How It Optimizes AI Agent Performance and Cost"?
HarnessPro iteratively changes an AI agent’s instructions, tools and runtime policies, then offers Lite, Medium and Heavy versions for different cost–performance trade offs.
What are the key points to validate first?
HarnessPro iteratively changes an AI agent’s instructions, tools and runtime policies, then offers Lite, Medium and Heavy versions for different cost–performance trade offs. Rather than optimize for one universal configuration, HarnessPro uses synthesized tests and Pareto based selection to retain options for lower cost, balance and higher performance.[3][11]
Sources
arxiv.orgHarnessPro: Recursive Harness Self-Improvement with ...
github.comHarnessPro/README.md at main · pzs19/HarnessPro · GitHub
pzs19.github.ioHarnessPro · Recursive Harness Self-Improvement
alirezashamsoshoara.github.ioDaily AI Papers — October 20, 2026
dev.toAgents Write the Commits Now — This Week's Research Is ...