# Harness² framework improves LLM agents on professional benchmarks

> Researchers introduce a recursive harness system that boosts performance of GLM‑5 and Gemini 3.5 Flash on three benchmark suites.

Oossa · 2026-10-14 · https://oossa.com/en/harness2-framework-improves-llm-agents-on-professional-benchmarks

A team led by Yanfei Chen released Harness², a plug‑and‑play framework that automatically refines the code that runs large language model agents. The system uses two layers of recursion: a task‑level step that proposes and tests edits without needing ground‑truth answers, and a harness‑level step that learns from those edits for future tasks. In tests on three professional benchmarks, a single recursion step lifted GLM‑5 and Gemini 3.5 Flash scores by up to 11.5 points on WorkBuddy‑Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB.

## The facts

- Paper posted on arXiv on 2026‑10‑14
- Improvement up to 11.8 percentage points on the LAB benchmark

## Why it matters

The method could let AI assistants adapt to new real‑world tasks without hand‑tuned code, benefiting developers who deploy LLM agents in changing environments.

## Sources & references

1. [Harness^2: Recursive Agent Harnessing for an Open World](https://arxiv.org/abs/2610.14224) – arXiv, 2026-10-14

Last updated: 2026-10-14