亚马逊云科技推出了两个面向 AI 智能体开发的开源项目：Strands Decider 2B 是一个用于在智能体工作流中作出决策的小型模型，Strands Box 则是一个用于控制智能体执行的沙箱。两个项目分别解决了构建能够使用工具、执行代码并与外部系统交互的智能体时所面临的不同问题。
Strands Decider 2B 是一个拥有 20 亿参数的模型，旨在从预定义选项中进行选择，而不是生成冗长的文本响应。它支持模型路由、工具选择、分类、护栏和上下文管理等任务。开发者无须让通用语言模型生成答案，而是可以提供一组选项，获得决策结果及相应的置信度分数。
该模型以 Qwen3.5-2B 为基础，将其语言模型头替换为一个用于为候选答案评分的指针头。它使用 rank-16 LoRA 适配器进行微调。亚马逊云科技已经发布了模型权重、训练数据和脚本，让开发者可以在本地运行并修改该模型。
据亚马逊云科技介绍，Strands Decider 2B 在 NVIDIA RTX 3090 上的决策延迟中位数约为 115 毫秒，在 Apple M3 MacBook 上处理小型任务时约为 153 毫秒。该公司还使用 JevBench 对模型进行了评估，其 v19 检查点在 231 项公开任务中的准确率为 72.3%。开发者可以安装 strands-decider 软件包，使用其命令行界面，或者将该模型集成到 Strands 智能体中。
第二个项目 Strands Box 目前处于开发者预览阶段，为执行 AI 智能体提供隔离环境。它将操作系统级隔离与 Dogwood 结合起来。Dogwood 是一门开源策略语言，用于定义智能体可以执行哪些操作。该沙箱支持 Shell 命令、Python 执行、MCP 工具、文件系统操作和网络访问。
Strands Box 使用由内核强制执行的隔离机制，将智能体交互路由至策略层。开发者可以根据操作内容和执行历史允许或拒绝相关操作。例如，策略可以要求智能体必须通过测试才能将代码推送到仓库，也可以限制外部 API 的调用频率。这些控制机制运行在语言模型之外，可以减少执行边界对提示词指令的依赖。
亚马逊云科技表示，Strands Box 可以配合不同的模型和智能体框架使用，无论它们运行在本地还是云环境中。这让开发者能够在不替换现有智能体架构的情况下引入执行策略。
Strands Decider 2B 的竞争对手包括 TypeSafe AI 的 Jev，后者同样是一款专门用于作出结构化决策，而不是生成冗长响应的模型。两者都支持分类、路由和工具选择，并会在返回决策结果的同时提供置信度分数。主要区别在于部署方式：Strands Decider 2B 是开源项目，可以在本地运行；Jev 则通过托管 API 提供。亚马逊云科技使用 JevBench 对 Strands Decider 进行了评估，但已公布的结果并不能证明它在性能上直接优于 Jev。
Strands Box 所解决的问题，也可以通过基于容器的沙箱、Firecracker microVM 和托管式智能体执行环境处理。传统沙箱主要限制文件系统、进程和网络访问，而 Strands Box 还增加了基于工具参数和历史操作的策略。这使开发者能够定义智能体在执行特定操作之前必须满足的条件。
早期开发者反馈主要关注将小型决策模型集成到智能体工作流中的实际限制。在一篇实践教程中，AWS Community Builder Yuuki Yamashita 表示：
Strands Decider 2B 可以阻止智能体调用不受支持的天气工具。
GitHub 讨论还发现了与并发请求、Apple Silicon 兼容性和冷启动延迟有关的问题，凸显了开发者在将该模型用于生产环境之前可能需要评估的方面。
AI Builder Dipen Vaja 在 X 上评论：
我宁愿让智能体在运行中途撞墙，也不愿让它沿着一条错误路径执行到底。亚马逊云科技今天开源了 Strands Box：这是一个内核级沙箱，将 Shell、Python、工具和网络访问都路由至同一个策略层，因此可以设置这样一条规则：智能体读取客户文件后，禁止其发起出站 HTTP 请求。
Strands Decider 2B 和 Strands Box 共同将 Strands 生态系统从智能体编排扩展至决策和执行控制领域。两个项目分别提供了用于选择智能体操作，以及限制这些操作如何执行的独立组件。
原文链接：
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Amazon Cloud Technology has launched two open-source projects for AI AI agent development: Strands Decider 2B, a small model for decision-making in AI agentworkflow, and Strands Box, a sandbox for controlling AI agent implementation. The two projects addressed the different problems faced in constructing AI agent, which could use tools, implement codes and interact with external systems.
Strands Decider 2B is a model with 2 billion parameters designed to choose from predefined options rather than generate lengthy text responses. It supports tasks such as model route, tool selection, classification, fences and context management. The developers do not need to allow the universal language model to generate answers, but can provide a set of options to obtain the decision results and corresponding confidence scores.
The model is based on Qwen3.5-2B and replaces its language model header with a pointer for scoring candidate answers. It fine-tunes with the rank-16 LoRA adapter. Amazon cloud technology has publish in place model weights, training data and scripts that allow developers to operate and modify the model locally.
According to the Amazon cloud technology presentation, the median decision delay of Strands Decider 2B on NVIDIA RTX 3090 was about 115 ms, and 153 ms for small tasks on Apple M3 MacBook. It also evaluated the model using JevBench, whose v19 inspection points had an accuracy of 72.3 per cent of 231 open missions. Developer can install the Strands-Decider package, use its command line interface, or integrate the model into Strands Z42Z.
The second project, Strands Box, is currently in the developer ' s preview phase, providing a quarantine environment for the implementation of AI AI agent. It combines operational system-level isolation with Dogwood. Dogwood is an open source policy language that defines what AI agent can do. The sandbox supports Shell command, Python execution, MCP tool, file system operation and network access.
Strands Box uses a kernel-enforced isolation mechanism to direct AI agent interactive circuits to strategic layers. The developer may allow or reject the operation on the basis of the content and history of the operation. For example, the strategy could require that AI agent must be tested in order to push the code to the warehouse or limit the frequency of calls from external API. These control mechanisms operate outside the language model and can reduce the dependence of the implementation of the border on the prompt directive.
Amazon cloud technology suggests that Strands Box can be used with different models and AI agent frames, whether they operate in local or cloud environments. This allows developers to introduce implementation strategies without replacing the existing AI agent architecture.
Strands Decider 2B ' s competitors include Jev of TypeSafe AI, who is also dedicated to structured decision-making rather than to the generation of long response models. Both support classification, route and tool selection and provide confidence scores while returning to decision-making outcomes. The main difference is deployment: Strands Decider 2B is an open source project that can be operated locally; Jev is provided through hosting API. The Amazon cloud technology used JevBech to evaluate Strands Decider, but the published results did not prove that it was directly superior to Jeff.
The problems addressed by Strands Box can also be handled through container-based sandboxes, Firecracker microVM and hosting AI agent. Traditional sandboxes mainly limit access to document systems, processes and networks, while Strands Box adds strategies based on tool parameters and historical operations. This enables developers to define the conditions that AI agent must meet before performing a particular operation.
Early developers ' feedback focused on the practical limitations of integrating small decision models into AI agentworkflow. In a practical course, AWS Community Builder Yuuki Yamashita says:
Strands Decider 2B prevents AI agent from calling unsupported weather tools.
The GitHub discussion also identified issues related to the simultaneous request, Apple Silicon compatibility and cold start delay, highlighting areas where developers might need to assess before using the model in the production environment.
Ii Builder Dipen Vaja commented on X:
I'd rather let AI agent hit the wall in the middle of the run than let it follow the wrong path to the end. Strands Box: This is a nuclear-grade sandbox that leads Shell, Python, tools and network access to the same policy level, so it is possible to set up a rule that AI agent, after reading the client files, is prohibited from initiating the HTTP request.
Strands Decider 2B and Strands Box jointly extended the Strands ecosystem from AI agent to the area of decision-making and implementation control. The two projects provided separate components for the selection of AI agent operations and for limiting how these operations were carried out.
Original link:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.