AI 模型和平台
代理AI基金会新增第一年达到10个项目
在 Google 上的首选来源中添加 Unite.AI
代理AI基金会宣布四个新主办项目于2026年10月22日：Agent技能，由Anthropic捐款；开放响应，由OpenAI捐款；OpenEnv，由Hugging Face捐款；OpenSandbox，由Alibaba捐款。 这些增加使非洲投资框架第一年的主办项目从2025年12月的三个发布项目增加到10个。
公告与2026年10月22日和23日在圣何塞麦凯内里会议中心举行的AGNTCON +MCCON北美赛事同时举行。 新项目与Agent2Agent(A2A)协议，Agent路由器，代理门道，AGENTS.md，鹅，模范背景协议(MCP)一起作为AAIF项目，在基金会的开放，供应商中立的治理下进行。 根据基础，这四个项目解决了代理堆栈的不同层：agents如何获得新技能，应用程序如何与模型跨供应商交谈，agents如何训练和评价，agents在哪里安全运行。
“一年前，AAIF 发布有3个项目，并且打赌代理AI需要一个中立的家才能兴旺。 今天，我们主办10,“称 Manik Surtani、CTO和代理AI基金会的联合创始人。 苏尔塔尼称四家不同的公司选择为中立治理提供项目，给社区一个建立共同标准、协议和基础设施的地方，没有一家公司控制结果。
Agent 技能
Agent技能是一种开放格式，用于赋予agents新的能力和域专业知识。 技能包指令，脚本，和资源以agents可按需加载的方式重复使用。 大卫·索里亚·帕拉(David Soria Parra)是MCP在Anthropic，称 Anthropic 发布 Agent技能的联合创造者和主要维护者，一年前，他给开发者和企业一个开放的标准，让agents更有能力，并且正在将其捐赠给AAIF，以便任何拥有技能的建筑都能够塑造接下来的东西。
在欢迎项目AAIF 称的基金会帖子中，规格是Apache 2.0和CC BY 4.0下的开源，超过70个agent平台已经支持它，超过100万个技能已经发布，安装了近2.8亿台。 文章指出，在7个月内，有100万种技能被列在技能.sh登记册上，这些技能用于软件工程、营销、设计、写作和研究。 基金会称 Agent技能将在AAIF下由技术指导委员会指导，由来自多个组织的核心维护者作为个人任职。
公开答复
Open Response是构建多提供者，互操作的LLM接口的开放规格，基于OpenAI Response API. 基础称它定义了共享计划，一致的流流事件，以及可扩展的工具，这样开发人员可以一次描述请求和输出，并运行在最小翻译的模型提供者之间。 OpenAI，称的技术人员史蒂夫·科菲(Steve Coffey)正在加入AAIF，以进一步推进其通过简单，统一的界面帮助开发者在语言模型上构建产品的任务。
规格上的一个基础帖子描述了通过API端点交换的请求和答复的通用格式，涵盖工具呼叫和流回。 规格允许扩展，由提供方特定名称确定，因此不同执行的特征并不冲突，而且由于它适用于API，采用它并不要求改变或再培训一种模式。 Databricks Model Service曝出Claude,Gemini,Databricks托管的开放响应端点，将请求翻译为每个提供者的本土格式，Vercel AI Gateway曝出OpenRespons API，为应用程序通过它的网关访问模型提供一个共同端点。
打开 Env
基金会将OpenEnv描述为创建强化学习环境的开放框架和社区枢纽，围绕一个集装箱化的服务构建，暴露了agent接口，MCP，以及训练图书馆使用的Gym训练协议。 Ben Burtenshaw是Hugging Face，称的机器学习工程师，OpenEnv背后的委员会正在分享其成员在Hugging Face,Microsoft,Reflection,Meta-PyTorch,Fleet,Mercor,RadixArk,Modal,Prime Intellect等地建造和使用的东西。
OpenEnv维护者的一个帖子描述了架构：一个客户端以重置()启动一集，以步骤()提交动作，并可以请求状态()，其结果包括一个观察，一个数字奖励，一集完成状态，以及元数据。 有些环境通过MCP暴露工具，允许外部agents以及培训期间使用这些工具。 文章重述图灵在OpenEnv上建立了Calendar Gym，团队报告任务命名明确日历标识符时任务成功率约为90%，而同一任务使用自然语言描述时任务成功率约为40%.
OpenEnv有一个多公司技术委员会和一个公开的AAIF建议，为共享环境规格寻找一个供应商中立的家。 维护者认为，路线图的优先事项包括将agent带与培训系统连接起来，验证环境，提供培训食谱，以及在培训期间支持任务取样、课程和评价，各个区域渔业委员会处于不同的审查和执行阶段，而不是即将发布的单一版本。
打开对话框
OpenSandbox是AI应用的通用沙盒平台，提供多种语言的SDK，统一的沙盒API，以及Docker和Kubernetes运行时间用于包括编码agents,GUI agents,agent评价，AI代码执行，以及强化学习培训在内的案例。 Alibaba开发基础设施负责人称 Xiaobin Xu。
描述项目的一个基础文章称OpenSandbox在Apache 2.0下获得许可，其沙盒协议是一套OpenAPI规格，主要分两部分：一个生命周期API用于创建，检查，挖掘，恢复，更新，快照和删除沙盒，一个执行API用于运行命令和代码，与文件合作，保持会话开放，并收集资源测量。 一个名为execed的执行守护进程运行在每个沙盒内并处理工作，Docker和Kubernetes运行时后端共享相同的生命周期接口，Docker后端直接创建和管理一个容器，Kubernetes后端生成工作量及其相关资源。
基金会背景
Linux基金会宣布 AAIF于2025年12月9日成立，创始捐款来自Anthropic的示范背景协议，Block的鹅，以及OpenAI的AGENTS.md. 组建时的白金会员有Amazon网络服务，Anthropic,Block,Bloomberg,Cloudflare,Google,Microsoft和OpenAI，以及Linux基金会称，当时MCP有超过10000个已公布服务器，AGENTS.md已经被超过60000个开源项目采用。 AAIF在代理AI中将自己描述为开放标准、协议和基础设施的中立之家。
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。
AI Models & Platforms
Agentic AI Foundation Adds Four Projects to Reach 10 in First Year
Add Unite.AI to your preferred sources on Google
The Agentic AI Foundation announced four new hosted projects on October 22, 2026: Agent Skills, contributed by Anthropic; Open Responses, contributed by OpenAI; OpenEnv, contributed by Hugging Face; and OpenSandbox, contributed by Alibaba. The additions bring AAIF to 10 hosted projects in its first year, up from three launch projects in December 2025.
The announcement coincides with AGNTCon + MCPCon North America, taking place October 22 and 23, 2026, at the San Jose McEnery Convention Center. The new projects join the Agent2Agent (A2A) protocol, Agent Router, agentgateway, AGENTS.md, goose, and the Model Context Protocol (MCP) as AAIF projects under the foundation’s open, vendor-neutral governance. According to the foundation, the four projects address different layers of the agentic stack: how agents gain new skills, how applications talk to models across providers, how agents are trained and evaluated, and where agents run safely.
“A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10,” said Manik Surtani, CTO and co-founder of the Agentic AI Foundation. Surtani said four different companies chose to contribute the projects to neutral governance, giving the community one place to build shared standards, protocols, and infrastructure with no single company controlling the outcome.
Agent Skills
Agent Skills is an open format for giving agents new capabilities and domain expertise. Skills package instructions, scripts, and resources in a reusable manner that agents can load on demand. David Soria Parra, co-creator and lead maintainer of MCP at Anthropic, said Anthropic launched Agent Skills a year ago to give developers and enterprises an open standard for making agents more capable, and is donating it to AAIF so that anyone building with skills can shape what comes next.
In a foundation post welcoming the project, AAIF said the specification is open source under Apache 2.0 and CC BY 4.0, that more than 70 agent platforms already support it, and that over 1 million skills have been published with nearly 280 million installations. The post reports that a million skills were listed on the skills.sh registry in seven months, with skills used across software engineering, marketing, design, writing, and research. The foundation said Agent Skills will be guided under AAIF by a Technical Steering Committee with core maintainers drawn from multiple organizations who serve as individuals.
Open Responses
Open Responses is an open specification for building multi-provider, interoperable LLM interfaces, based on the OpenAI Responses API. The foundation said it defines a shared schema, consistent streaming events, and extensible tooling so developers can describe requests and outputs once and run them across model providers with minimal translation. Steve Coffey, a member of technical staff at OpenAI, said the project is joining AAIF to further its mission of helping developers build products on language models through simple, unified interfaces.
A foundation post on the specification describes a common format for the requests and responses exchanged through an API endpoint, covering tool calls and streaming responses. The specification permits extensions, identified by provider-specific names so features from different implementations do not conflict, and because it applies to the API, adopting it does not require changing or retraining a model. The post identifies two existing implementations: Databricks Model Serving exposes an Open Responses endpoint for Claude, Gemini, and Databricks-hosted open models, translating requests into each provider’s native format, and Vercel AI Gateway exposes an OpenResponses API that gives applications a common endpoint for accessing models through its gateway.
OpenEnv
The foundation describes OpenEnv as an open framework and community hub for creating reinforcement learning environments, built around a containerized service that exposes an agent interface, MCP, and the Gym training protocol used by training libraries. Ben Burtenshaw, a machine learning engineer at Hugging Face, said the committee behind OpenEnv is sharing what its members have built and used at Hugging Face, Microsoft, Reflection, Meta-PyTorch, Fleet, Mercor, RadixArk, Modal, Prime Intellect, and more.
A post by the OpenEnv maintainers describes the architecture: a client starts an episode with reset(), submits actions with step(), and can request state(), with results that include an observation, a numeric reward, episode completion status, and metadata. Some environments expose tools through MCP, allowing those tools to be used by external agents as well as during training. The post recounts that Turing built Calendar Gym on OpenEnv, and the team reported roughly 90% task success when tasks named explicit calendar identifiers, compared with roughly 40% when the same tasks used natural-language descriptions.
OpenEnv has a multi-company Technical Committee and a public AAIF proposal seeking a vendor-neutral home for the shared environment specification. According to the maintainers, roadmap priorities include connecting agent harnesses to training systems, validating environments, providing training recipes, and supporting task sampling, curriculum, and evaluation during training, with individual RFCs at different stages of review and implementation rather than a single upcoming release.
OpenSandbox
OpenSandbox is a general-purpose sandbox platform for AI applications, offering multi-language SDKs, unified sandbox APIs, and Docker and Kubernetes runtimes for use cases including coding agents, GUI agents, agent evaluation, AI code execution, and reinforcement learning training. “While OpenSandbox was created at Alibaba, we don’t want its future to be defined by a single company,” said Xiaobin Xu, head of development infrastructure at Alibaba.
A foundation post describing the project states that OpenSandbox is licensed under Apache 2.0 and that its sandbox protocol is a set of OpenAPI specifications with two main parts: a lifecycle API for creating, inspecting, pausing, resuming, renewing, snapshotting, and deleting a sandbox, and an execution API for running commands and code, working with files, keeping sessions open, and collecting resource measurements. An execution daemon named execd runs inside each sandbox and handles that work, and the Docker and Kubernetes runtime backends share the same lifecycle interface, with the Docker backend creating and managing a container directly and the Kubernetes backend creating a workload and its related resources.
Foundation Background
The Linux Foundation announced AAIF’s formation on December 9, 2025, with founding contributions of Anthropic’s Model Context Protocol, Block’s goose, and OpenAI’s AGENTS.md. Platinum members at formation were Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI, and the Linux Foundation said at the time that MCP had more than 10,000 published servers and that AGENTS.md had been adopted by more than 60,000 open source projects. AAIF describes itself as the neutral home for open standards, protocols, and infrastructure in agentic AI.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.