HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs

HarnessPro. Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs. By Zhuoshi Pan and eight others, from Tsinghua University, Z.AI, and three other institutions. Published on the nineteenth of October. Its code repository is worked out: about two hundred source files, with a readme and a licence.

Harness development plays a crucial role in building capable AI agents, and recent approaches improve harnesses through recursive self-improvement, or RSI, letting the system revise its own harness over repeated rounds. The trouble is the bill. Their performance gains often come with higher execution costs. A harness that solves more tasks while spending more on every one of them is a trade, and not every deployment can afford the expensive side of it.

HarnessPro is a label-free method for evolving harnesses that aims to improve task performance and reduce execution cost together. Label-free means candidate harnesses are judged without ground-truth solutions and without official tests: they are evaluated using tests generated by an LLM instead. And rather than converging on a single winner, it keeps a spread of harnesses that strike the balance between performance and cost differently.

Each iteration is carried out by specialized agents. Some diagnose execution trajectories. Others plan updates under different performance-cost priorities, and those plans are implemented as candidate harnesses. Selection uses normalized weighted squared-distance selection to return three harnesses rather than one: a cost-focused lite version, a balanced medium version, and a performance-focused heavy version, each meant for a different deployment budget. The selected harnesses enter the next iteration, where execution feedback guides further updates.

Experiments cover Terminal-Bench 2, SWE-Bench Pro and ProgramBench, under two pairings: GPT-5.5 with Codex, and GLM-5.2 with Claude Code. With GPT-5.5 and Codex, the heavy mode of HarnessPro achieves an average pass-rate gain of two percentage points across these benchmarks. Alongside it comes an average per-task cost reduction of twenty-six point one percent. Both are measured relative to RHO. The advantage also holds with GLM-5.2 and Claude Code.

The result runs against the pattern described at the start. Earlier self-improvement tended to pay for performance with cost, but here even the most performance-focused of the three harnesses scores higher and costs less per task than the comparison. The lite and medium modes then cover budgets where cost weighs even more, so the choice of trade-off is handed to whoever deploys the agent.