Google开源Harness2递归Agent框架
Google Open-Sources Harness2, a Recursive Agent Harnessing Framework
Google Cloud AI Research has open-sourced Harness². It is a recursive agent harnessing framework that improves the code and files around a frozen model. For developers, the pitch is simple. Your agent may get better without a better model.
Google Cloud AI Research 已开源 Harness²。这是一个递归智能体框架(agent harnessing framework),用于改进围绕冻结模型的代码和文件。对开发者而言,卖点很简单:即使模型没有变好,你的智能体也可能变得更好。
TL;DR
TL;DR(太长不看版)
- Size: A framework, not a model. 0 parameters trained; 8 editable harness components. Version 0.1.0.
- Runs on: Linux with Python 3.12 or 3.13 and uv. Model runs call models through Vertex AI. GPU requirement: Not disclosed.
- Performance: 1 refinement step improved all 4 model and harness setups on all 3 benchmarks.
- Best: GLM-5 under OpenCode rose from 64.2 to 75.7 on WorkBuddy-Bench (+11.5 points).
- Worst: GLM-5 under Codex gained only 1.2 points on JobBench (23.6 to 24.8).
- 规模:是一个框架,而非模型。训练了 0 个参数;包含 8 个可编辑的框架组件。版本 0.1.0。
- 运行环境:Linux,Python 3.12 或 3.13 以及 uv。模型运行通过 Vertex AI 调用模型。GPU 要求:未公开。
- 性能:经过 1 次优化步骤后,在所有 3 项基准测试中,所有 4 种模型和框架组合均得到提升。
- 最佳表现:在 OpenCode 下,GLM-5 在 WorkBuddy-Bench 上的得分从 64.2 提升至 75.7(+11.5 分)。
- 最差表现:在 Codex 下,GLM-5 在 JobBench 上仅获得 1.2 分的提升(从 23.6 到 24.8)。
What is Harness²?
什么是 Harness²?
Harness² is a plug-and-play framework from Google Research that refines an agent’s harness during deployment. A harness is everything around the model: prompts, tools, memory, skills and orchestration. Harnessing is the process of optimizing that system. Model weights never change.
Harness² 是 Google Research 推出的即插即用框架,可在部署期间优化智能体的框架(harness)。框架指的是模型周围的一切:提示词、工具、记忆、技能和编排。优化(Harnessing)是指对该系统进行优化的过程。模型权重永不改变。
The target setting is the open world. Tasks arrive unseen, and no verifier says whether an answer is right. The research team argues that most self-improvement methods assume labels or test cases. Harness² is built to work without them.
目标场景是开放世界。任务以未见过的形式出现,且没有验证器来判定答案是否正确。研究团队指出,大多数自我改进方法都假设存在标签或测试用例。Harness² 的设计旨在无需这些条件即可工作。
What are the 8 harness components?
那 8 个框架组件是什么?
Harness² exposes 8 editable components, each mounted as a file:
Harness² 暴露了 8 个可编辑组件,每个组件都挂载为一个文件:
- System prompt
- Guardrails
- Memory
- Tool guidance
- Subagents
- Skills
- Scripts
- Plugins (middleware)
- 系统提示词(System prompt)
- 护栏(Guardrails)
- 记忆(Memory)
- 工具引导(Tool guidance)
- 子智能体(Subagents)
- 技能(Skills)
- 脚本(Scripts)
- 插件(中间件)(Plugins (middleware))
They cover context construction, the tool interface, delegation and behavior inside the tool loop. The layout extends the 7-file split from earlier Agentic Harness Engineering work by 1 component.
它们涵盖了上下文构建、工具接口、委托以及工具循环内的行为。该布局在早期 Agentic Harness Engineering 工作的 7 文件划分基础上增加了 1 个组件。
How does recursive agent harnessing work?
递归智能体框架优化是如何工作的?
The framework runs two linked levels of recursion:
该框架运行两个相互关联的递归层级:
- Task-level recursion uses a contrastive Propose, Probe, Compose step. First, the agent runs the task once with the base harness. An editor then proposes 2 candidates. The invariant candidate targets general agent failures such as planning or tool use. The domain candidate targets knowledge and procedures for the current domain. Both candidates run on the same task. The composer compares their trajectories against the reference and keeps only edits backed by observed behavior.
- Harnessing-level recursion happens after each task. A reflector stores the adopted harness plus an EXPERIENCE.md file of lessons. This library guides future proposals. Each new task still starts from the base harness.
- 任务级递归使用对比式的提议(Propose)、探测(Probe)、组合(Compose)步骤。首先,智能体使用基础框架运行一次任务。然后,编辑器提出 2 个候选方案。不变候选方案针对通用智能体失败情况,如规划或工具使用。领域候选方案针对当前领域的知识和流程。两个候选方案都在同一任务上运行。组合器将它们的轨迹与参考轨迹进行比较,仅保留有观察行为支持的编辑。
- Harnessing-level recursion 发生在每个任务之后。反射器存储已采用的 harness 以及包含经验教训的 EXPERIENCE.md 文件。该库指导未来的提案。每个新任务仍然从基础 harness 开始。
The budget k scales the search. Parallel mode probes k invariant and k domain candidates, then makes k compositions. Sequential mode chains k refinement steps.
预算 k 缩放搜索。并行模式探测 k 个不变量和 k 个域候选项,然后进行 k 次组合。顺序模式链接 k 个细化步骤。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力