AI Companies Compete Over Agent Harnesses
36Kr reports that AI companies are turning agent runtime harnesses into hosted services. Microsoft expanded Copilot Harness across its products and offered managed services through the cloud, while Moonshot AI released Kimi Hosted Agents with a “Harness as a Service” offering. OpenAI and Anthropic had also made their internal runtime systems available to developers, while DeepSeek open-sourced its harness framework. In a Microsoft technical white paper, Microsoft reported that keeping Claude Opus 4.8 unchanged but switching the harness behind Copilot Studio increased the complex-task pass rate by 22.1 percentage points. The new system nearly doubled model calls per task, yet reduced calculated input-token consumption by 21%.
Harnesses coordinate tool use, multi-step execution and task state, and their design can affect task success, efficiency and operating costs. The report says their availability as reusable services can influence enterprise deployment and maintenance choices and give providers visibility into real task execution, subject to data-rights limits.
11月7日,微软发布了一份关于Agent执行系统的技术白皮书,包含一组有点出人意料的测试结果:同样使用Claude Opus 4.8,只换掉Copilot Studio背后的Harness,复杂任务通过率就提高了22.1个百分点。新系统每项任务调用模型的次数几乎翻倍,折算后的输入Token消耗却可以减少21%。
就在十天前,微软刚宣布将统一的Copilot Harness推广到旗下多个产品,并通过云平台向开发者出售托管服务。
11月5日,月之暗面刚刚发布Kimi Hosted Agents(托管智能体),正式提出Harness as a Service。8月,DeepSeek开源了自己的Harness;OpenAI、Anthropic更早就已经把内部运行系统变成开发者可以调用的服务。
那个大模型背后的“运行系统”,开始成为模型公司眼中的关键先生。
#Microsoft#Anthropic#OpenAI#Moonshot AI#DeepSeek#Copilot Studio#Copilot Harness#Claude Opus 4.8#Kimi Hosted Agents#Managed Agents#Agents API#DeepSeek Harness