Google Research 发布 Harness² 框架，通过递归优化提升智能体性能

Google Research 推出了 Harness²，一个用于递归智能体优化的即插即用框架1。该框架在不修改模型权重的前提下，通过任务级和知识蕴含级两层递归机制优化智能体系统1。框架支持优化的可编辑组件包括系统提示词、防护栏、记忆、工具指导、子智能体、技能、脚本和插件共八个方面1。

Harness² 在三个专业基准测试中取得显著效果提升1。在 WorkBuddy-Bench 上提升 11.5 分，JobBench 上提升 4.5 分，LAB 部分通过率提升 11.8 个百分点1。框架已验证支持 GLM-5 和 Gemini 3.5 Flash 两种模型1，并采用 Apache 2.0 开源许可证发布1。框架支持平行和顺序递归两种工作模式1。

Google Research has introduced Harness², a plug-and-play framework designed to optimize recursive intelligent agent systems without modifying model weights 1. The framework operates through two-layer recursion mechanisms at both the task and knowledge-embedding levels, enabling optimization of eight editable agent components including system prompts, guardrails, memory, tool guidance, sub-agents, skills, scripts, and plugins 1.

Performance evaluations across three professional benchmarks demonstrated substantial improvements 1. On WorkBuddy-Bench, the framework achieved an 11.5-point gain, while JobBench saw a 4.5-point improvement, and LAB recorded an 11.8 percentage point increase in pass rate 1. The optimization benefits applied to both GLM-5 and Gemini 3.5 Flash models 1. Harness² supports two operational modes—parallel and sequential recursion—providing flexibility in how recursive optimization is applied 1.

The framework has been released under the Apache 2.0 open-source license, with the corresponding research paper published in 2026 and available as an arXiv preprint under number 2610.14224 1.