Cisco Outshift Launches Open-Source Project for AI Agent Protocols

Cisco Outshift 推出 AI 智能体协议开源项目 传统 REST API 在非确定性多智能体工作流中失效，需要涵盖消息传递、身份与发现的全新开放标准，以防供应商锁定。 Tushar Agrawal 概述了从传统确定性软件向非确定性 AI 智能体交互的转变，并介绍了专为智能体工作流设计的改良版 OSI 模型。他展示了由 Cisco、Dell、Google、Oracle 和 Red Hat 支持的 Linux 基金会开源倡议 AGNTCY，旨在建立厂商中立的协议规范。该架构基于四大核心支柱：智能体目录（OASF）、智能体身份、SLIM 消息传递以及多智能体可观测性。本次演讲最后现场演示了多智能体咖啡供应链场景，展示了发布/订阅消息传递、MCP 集成和基于工具的访问控制。 思科旗下孵化部门 Outshift 提出“智能体互联网”（Internet of Agents）架构，倡导超越脆弱的硬编码 API，构建适用于多智能体自主协作的新型 OSI 模型。 (4:31) 思科联合戴尔、谷歌、甲骨文、红帽及超过 75 家初创公司，在 Linux 基金会旗下发起开源项目 AGNTCY，旨在构建中立的标准生态以防止厂商锁定。 (9:38) AGNTCY 围绕四大核心支柱构建：基于开放智能体模式框架（OASF）的去中心化智能体目录、跨 IdP 的智能体身份认证、基于 gRPC 的安全低延迟交互消息协议（SLIM），以及跨平台的多智能体可观测性。 (11:23) 现场演示展示了咖啡供应链的多智能体协作，结合了 A2A 通信、MCP 工具调用与发布/订阅模式，并成功展现了针对不同智能体的细粒度工具访问控制与身份鉴权。 (19:02)

Tushar Agrawal of Cisco's Outshift argues that scaling autonomous AI agents requires moving beyond rigid, deterministic APIs to an open, standardized 'Internet of Agents' that enables dynamic, secure collaboration across cross-cloud and multi-vendor environments. Moving Beyond Deterministic Architectures (1:57) Agrawal asserts that transitioning from traditional cloud computing to agentic software demands a fundamental architectural shift. Traditional architectures such as virtual machines and microservices rely on structured, deterministic request-response paths. In contrast, autonomous AI agents operate non-deterministically, requiring systems to dynamically interpret context, reason through changing parameters mid-task, course-correct, and react to actions executed by other agents. Because real-world user workflows span fragmented software ecosystems, unlocking the full utility of agents requires autonomous collaboration across heterogeneous environments. Rather than depending on hardcoded and brittle API integrations, agents must interoperate seamlessly across differing underlying frameworks, hosting clouds, Identity Providers (IdPs), and operational personas. The Internet of Agents and AGNTCY Initiative (4:31) To bridge human-like reasoning with software-scale execution, Cisco advocates for an 'Internet of Agents' that expands traditional networking abstractions. Agrawal compares multi-agent interaction to children building a toy house together: each participant must observe adjustments and adapt their next move in real time. Standard REST APIs lack the flexibility for this dynamic exchange, which requires conversational, low-latency interfaces capable of handling interruptions and multi-party dialogue. Architecturally, this model introduces syntactic interoperability layers (such as Agent-to-Agent protocols and Model Context Protocol) and semantic layers for dynamic task decomposition and policy planning. To prevent proprietary vendor lock-in, Cisco initiated AGNTCY as an open-source Linux Foundation project. Founding corporate members include Outshift by Cisco, Dell, Google, Oracle, and Red Hat, alongside a consortium of over 75 participating startups and contributors. Four Core Pillars of AGNTCY Architecture (11:23) The AGNTCY technical stack is built on four core pillars: an Agent Directory, Agent Identity, Secure Low-Latency Interactive Messaging (SLIM), and Multi-Agent Observability. The Agent Directory utilizes the Open Agentic Schema Framework (OASF) as a decentralized registry where agents publish capabilities and invocation schemas for autonomous discovery. The Identity layer manages cross-IdP trust, ensuring task-level permissions and governance policies are strictly enforced when agents communicate across organizational borders. For transport, SLIM extends gRPC into an encrypted publish-subscribe protocol, minimizing latency overhead for Agent-to-Agent (A2A) and Model Context Protocol (MCP) integrations. Completing the lifecycle across discovery, composition, deployment, and evaluation, the observability pillar standardizes tracing and telemetry to provide unified visibility across multi-agent executions. Validating Governance in a Multi-Agent Simulation (19:02) To demonstrate the framework in operation, Agrawal presented a simulated coffee supply chain featuring a buyer agent and three regional supplier agents representing Brazil, Colombia, and Vietnam built with LangGraph. When the buyer agent broadcast an open inventory inquiry, SLIM handled the interaction via a many-to-many pub/sub topology, while an MCP connection allowed the Colombia agent to incorporate live local weather data directly into its inventory evaluation. The demonstration also highlighted tool-based access controls during direct request-response transactions. When the buyer submitted a formal purchase order to the Colombia agent, authenticated identity validation allowed the transaction to proceed. However, when the exact same order was directed to the Brazil agent, the system blocked execution because Brazil's identity lacked verified authorization for transactional tasks, proving that granular access policies can restrict high-risk operations while permitting general informational queries. Agrawal concludes by emphasizing that vendor-neutral governance is essential for agentic interoperability, inviting developers and enterprise architects to contribute directly to AGNTCY's working groups focused on identity, messaging transport, telemetry observability, and agent discovery.

Cisco 旗下新兴技术孵化部门 Outshift 的生成式 AI 产品管理高级总监 Tushar Agrawal 指出，人工智能正从传统的确定性软件向自主推理的非确定性智能体演进，行业亟需建立一套开放、互操作且中立的“智能体互联网”（Internet of Agents）架构标准，以防止生态陷入厂商绑定。 从确定性软件向非确定性智能体的范式转变 (1:57) Agrawal 认为，传统基于虚拟机（VM）与微服务架构的云计算基础设施依赖严格固定的请求-响应路径，本质上属于确定性工作流。与之相对，新兴 AI 智能体具备自主行动力与非确定性行为特征，在执行任务时需要动态理解语境、自主修正推理路径，并对其他智能体的行为做出灵活反应。随着企业日常流程依赖大量异构软件工具，多智能体协作若仅依赖容易断裂的硬编码 REST API，将难以应对跨云环境、不同框架及多重身份提供商（IdP）带来的复杂变量。 为弥合软件规模化执行与类人推理之间的鸿沟，Cisco 提倡借鉴早期互联网网络协议及 DNS 域名解析的标准化经验，重构面向智能体时代的开放系统互连（OSI）分层模型。该框架在传统网络层之上扩展了语法互操作层与语义智能体层：语法层依托 A2A（智能体对智能体）协议与 MCP（模型上下文协议）规范交互格式；语义层则承担任务分解、策略对齐与意图规划。Agrawal 以儿童共同搭建积木模型为喻，强调多智能体系统必须依赖低延迟且支持打断、澄清与多轮迭代的会话型接口，以替代死板的传统接口形式。 开源倡议 AGNTCY 与互操作体系四大技术支柱 (9:38) 为了在跨厂商环境中推行统一标准并规避生态壁垒，Cisco 联合 Dell、Google、Oracle、Red Hat 等创始成员，以及超过 75 家初创公司与开源贡献者，在 Linux 基金会旗下发起了名为 AGNTCY 的开源项目。该项目旨在建立中立的技术标准，覆盖智能体全生命周期的四个阶段，即发现（Discover）、组合（Compose）、部署（Deploy）与评估（Evaluate）。 AGNTCY 架构设立了四个核心技术支柱以支撑跨云协作：一是基于开放智能体模式框架（OASF）的“智能体目录”（Agent Directory），作为去中心化注册表发布智能体能力与调用模式以实现动态发现；二是“智能体身份”（Agent Identity），用于解决跨 IdP 的信任凭证验证并界定任务级策略权限；三是“SLIM 消息传输机制”（Secure Low-Latency Interactive Messaging），该协议基于 gRPC 拓展出内置加密能力的高性能发布-订阅总线，降低 A2A 与 MCP 通信的延迟开销；四是“多智能体可观测性”（Multi-Agent Observability），为异构智能体平台提供跨系统的遥测追踪与评估标准。 多拓扑通信与工具集成的供应链实测 (19:02) 在模拟咖啡农场采购的演示中，演讲者展示了一个由四个 LangGraph 智能体组成的网络：一个买方智能体与三个分别代表巴西、哥伦比亚和越南的供应方智能体，底层依托基于 SLIM 的 A2A 协议实现安全通信。同时，哥伦比亚智能体通过 MCP 服务直接对接外部实时天气应用，使气象环境数据能够实时参与供应状况评估。 在通信拓扑上，SLIM 既支持广播也支持点对点交互。当买方智能体发布“最新全球咖啡库存”这一全局查询时，SLIM 采用多对多发布-订阅（Pub/Sub）模式将消息同时推送给所有订阅该主题的供应方智能体。在此阶段，即便巴西智能体的身份尚未取得交易级授权，安全策略仍允许其返回基础信息；哥伦比亚智能体则同步调用 MCP 天气数据，各方协同向买方智能体返回整合后的库存信息。 基于身份验证的细粒度工具访问控制机制 (21:05) 在完成全局库存检索后，买方智能体转向点对点请求-响应模式以执行特定采购任务，提出“以每磅 5 美元购买 500 磅哥伦比亚咖啡”的具体订单。SLIM 将该请求精准定向路由至哥伦比亚智能体，底层系统即时核验其认证身份并校验交易权限范围，最终成功执行并完成了该笔交易订单。 为了验证系统内嵌的治理约束，现场随后发起了向巴西智能体采购相同货物的尝试。尽管巴西智能体在前序信息交互中表现正常，但由于其未取得交易执行的身份验证资质，系统触发基于工具的访问控制策略（Tool-Based Access Control），安全层直接阻断了该操作并返回权限受限错误。演讲者指出，这种细粒度权限管控证明了在开放的跨组织智能体交互中建立强鉴权机制的可行性与必要性。 Agrawal 呼吁行业开发者与企业共同加入 Linux 基金会旗下 AGNTCY 的身份、传输、可观测性与目录等工作组。尽管该项目提供了跨云智能体协作的原型参考，但在异构企业生态中，跨身份提供商的大规模信任链落地与通用动态治理规范的最终确立，仍取决于多厂商长期的开源协同与行业采纳深度。

## Background 本视频由 Ai4 频道于 2025 年 8 月 24 日发布在 YouTube 平台上。视频记录了 Cisco 旗下 Outshift 团队 GenAI 产品管理高级总监 Tushar Agrawal 的主题演讲，主要探讨了多智能体软件协作的发展趋势，并介绍了 AGNTCY 倡议为“智能体互联网（Internet of Agents）”所构建的基础技术与开放标准。 ## TL;DR 演讲者阐述了计算范式从确定性微服务向非确定性 AI 智能体跨越的过程，并指出跨厂商多智能体协作需要全新的网络与协议标准。为了应对智能体间的发现、身份认证、安全通信与可观测性等挑战，Cisco 联合开源社区推出了 AGNTCY 倡议以及 OASF、SLIM 等核心技术规范。最后，演讲通过咖啡供应链的多智能体协作案例，实操演示了发布订阅通信、工具级权限控制以及端到端链路追踪等关键特性。 ## Atmosphere & style 整场演讲风格专业且极具前瞻性，呈现出浓厚的技术探研与开源分享氛围。主讲人以沉稳干练的语气循序渐进地剖析复杂架构，并通过生动的现实比喻与咖啡供应链 Demo 将抽象的协议概念具象化。现场互动诚恳真挚，展现出拥抱跨平台互操作性与推进行业共同标准的开放姿态。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=4s) 随着 AI 软件开发从单模型助手转向多智能体协同，计算模式正在从传统的确定性计算向具备自主性和上下文感知能力的非确定性 AI 智能体跨越。 - [03:14](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=194s) 未来的智能体生态将高度碎片化，智能体由不同厂商开发、部署在不同云平台且使用不同的身份提供商（IDP），因此迫切需要构建统一的“智能体互联网”（Internet of Agents）。 - [03:14](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=194s) 智能体之间的沟通更加“交谈化”（Chatty），具备打断、纠偏和协商能力，超越了传统 API 固定的“请求-响应”模式，需要类似 A2A 和 Agent Connect Protocol 等新协议。 - [07:46](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=466s) 为了支持智能体协同，传统 OSI 模型需要在应用层之上扩展出“语法层”（如 MCP、A2A通信标准）与“语义层”（任务分解与内置权限规划）。 - [07:46](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=466s) 思科（Cisco）发起了开源倡议 AGNTCY，现已被 Linux 基金会托管，由 Cisco Outshift、Dell、Google、Oracle 和 Red Hat 等团队共同发起，已有超过 75 家初创公司加入。 - [11:23](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=683s) AGNTCY 包含了四大核心组件：基于 OASF 的智能体目录（Agent Directory）、跨 IDP 的智能体身份认证（Agent Identity）、扩展 gRPC 的 Slim 低延迟消息协议（Messaging），以及跨平台统一的可观测性（Observability）。 - [15:43](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=943s) 智能体生命周期包含四大架构阶段：发现（Discover）、组合（Compose）、部署（Deploy）和评估（Evaluate）。 - [19:02](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1142s) 在咖啡供应链（Coffee Farm）多智能体演示中，各个智能体基于 LangGraph 构建，利用 A2A 进行交互，并通过 MCP 接入外部天气 API 数据。 - [19:02](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1142s) 演示展示了基于 SLIM 消息层的广播式“多对多发布订阅模式”（Many-to-Many Pub/Sub），实现跨智能体实时广播与分布式数据汇总。 - [21:06](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1266s) 演示验证了基于智能体身份的工具级访问控制（Tool-based Access Control），具备权限的智能体可以成功下单，而未经授权的智能体会被直接拒绝执行敏感操作。 - [22:15](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1335s) AGNTCY 设立了涵盖身份识别（Identity）、通信网格（SLIM）、可观测性（Observe）和目录服务（Directory）等专项工作组，呼吁行业共同推进开放互操作标准。 ## Detailed walkthrough 以下是根据视频转录内容（第一部分）整理的详细且忠实于原意的文字记录： ### [00:04](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=4s) 出发点与背景：确定性计算向非确定性 AI 智能体的跨越 演讲由来自 Cisco 孵化团队 Outshift 的 GenAI 产品管理高级总监 Tushar Agrawal 主讲。Tushar 首先表达了对与会者的感谢，并简要介绍了自己的背景：拥有约 15 年跨大型企业和初创公司的产品管理经验，其 AI 旅程始于 AWS，当时他负责计算机视觉系列服务的产品管理。 随后，Tushar 介绍了 Outshift 的定位——它是 Cisco 内部的创新孵化部门，旨在探索前沿新兴技术，为 Cisco 的未来发展探索新领地，并将这些创新成果整合进现有的或全新的产品组合中，以助力客户成功。 在引入主题时，Tushar 指出，当前需要以一种全新的视角来看待智能体（Agent）之间的协作与互操作性。这种模式与过去微服务（Microservices）和虚拟机（VMs）之间的协作有着本质的不同： * **确定性与非确定性的转变**：传统软件协作依赖于确定的、硬编码的格式；而 AI 智能体拥有自主性（Agency），要求它们能够根据上下文实时理解、反应并动态调整与其它智能体的协作方式。 * **云计算新纪元**：当前的云计算大部分仍处于确定性计算时代，而智能体协作标志着迈入非确定性时代，必须重新构想底层的基础架构。 --- ### [03:14](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=194s) 智能体互联网（Internet of Agents）的概念与挑战 从用户的实际工作场景来看，完成一项复杂任务通常需要跨越多个不同的工具和应用程序，因此不可能通过硬编码或确定性流程来实现。智能体之间必然需要进行深度的协同合作。 **跨厂商协作的复杂性** 未来的智能体生态将高度碎片化。这些智能体将由不同的供应商（Vendors）开发、基于不同的开发框架建立、部署在不同的云平台上、使用不同的身份提供商（IDP），并归属于不同的企业组织。它们有着各不相同的属性与角色，需要根据具体的任务要求进行动态调整。 **构建“智能体互联网”（Internet of Agents）** 为了解决这一挑战，Tushar 提出了“智能体互联网”的构想。这一概念旨在融合软件的特性（高速度、大规模扩展）与人类的特性（动态性、自主性、推理能力和任务理解能力）。 Cisco 曾为 Web 时代的互联网确立了标准的网络协议栈（如 DNS 域名解析），而在智能体时代，同样需要一套全新的标准和协议，以实现智能体之间的**发现（Discover）、身份认证（Identify）、信任建立（Trust）以及政策合规（Policy Adherence）**。 为了直观形象地解释这种协作方式，Tushar 提到他在 Perplexity 上搜索过一个生动的比喻： > 想象四个孩子正在一起用积木搭房子。如果一个孩子在某个位置放了一块积木，另一个孩子就必须对此作出反应，并在另一个位置放置相应的积木。 传统的 API 接口缺乏这种灵活性——API 主要是固定的“请求-响应”（Request-Response）模式；而智能体之间的沟通则更加“交谈化”（Chatty），需要具备打断（Interrupt）、纠偏（Course Correct）和进一步澄清（Clarify）的能力。这超越了现有 API 的能力范畴，也是类似 A2A（Agent-to-Agent）或 Agent Connect Protocol 等新协议致力于解决的问题。此外，在多厂商、多环境的复杂体系中，还需要具备统一的端到端可观测性（Observability），形成“观察-评估-迭代”的闭环。 ``` 传统 API 模式: [客户端] ------ 请求 ------> [服务端] [客户端] <----- 响应 ------ [服务端] 智能体协作模式: [智能体 A] <--- 交互/打断/纠偏 ---> [智能体 B] \ / <--- 动态协商/上下文 ---> ``` --- ### [07:46](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=466s) OSI 模型的演进与 AGNTCY 开放开源倡议 为了应对这些技术挑战，传统用于网络通信的 OSI 七层模型（底层为物理层、数据链路层、网络层、传输层，上层为会话层、表示层、应用层）需要被赋予新的内涵。如果说互联网时代关注的是虚拟机（VMs），云计算时代关注的是微服务（Microservices），那么智能体时代则需要在应用层之上建立新的**语法层（Syntactic Layer）**和**语义层（Semantic Layer）**。 * **语法层**：处理如 MCP（Model Context Protocol）和 A2A 等基础通信标准。 * **语义层**：理解任务本身的含义。例如，将用户的请求分解为具体的执行计划，并将所需的权限控制（Permissions）直接内置于规划之中，而非事后补救。 ``` +-------------------------------------------------------+ | 语义层 (Semantic): 任务分解、权限规划、智能体推理 | +-------------------------------------------------------+ | 语法层 (Syntactic): MCP、A2A、协议互操作性 | +-------------------------------------------------------+ | 经典应用层 (Application / Session / Presentation) | +-------------------------------------------------------+ | 基础网络层 (Physical / Data Link / Network / Trans) | +-------------------------------------------------------+ ``` **AGNTCY 倡议与其生态建立** 为了将“智能体互联网”变为现实，Cisco 在今年早些时候推出了名为 **AGNTCY**（故意去掉了字母 e）的倡议。目前，该项目已正式成为 **Linux 基金会（Linux Foundation）** 旗下的开源项目。 - **参与规模**：已有超过 75 家初创公司加入，共同构建跨厂商的多智能体协作框架与标准。 - **发起成员**：除了 Cisco 旗下的 Outshift 外，创始成员还包括来自 **Dell、Google、Oracle 和 Red Hat** 的团队。 - **行业呼吁**：演讲者现场呼吁行业开发者与企业积极加入这一开源倡议，共同打造具有跨平台互操作性的智能体互联网。 --- ### [11:23](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=683s) AGNTCY 架构四大核心组件 针对多厂商智能体协作中的核心痛点，AGNTCY 框架定义了四大技术组件： #### 1. 智能体目录（Agent Directory）与 OASF 当智能体或编排器（Orchestrator）接收到非预先设定的任务时，首先需要寻找具备相应能力的其它智能体。为此，AGNTCY 推出了 **OASF（Open Agentic Schema Framework，开放智能体模式框架）**，用于以标准化方式描述智能体的能力、调用接口及配置参数。Agent Directory 以 OASF 为基础，提供去中心化的智能体注册与同步机制，使企业内部及外部（第三方供应商或其它企业）的智能体能够安全地互相发现与验证。 #### 2. 智能体身份认证（Agent Identity） 解决跨组织/跨身份提供商（IDP）的权限与安全问题。当使用不同 IDP 的智能体相互交互时，系统需要精准判定并限制其可执行的具体任务范围。 #### 3. 智能体消息传递（Messaging）与 Slim 协议 在包含多个智能体的集群（Ensemble）中，传统的通信开销极大。AGNTCY 扩展了 **gRPC**，推出了 **Slim（Secure Low-Latency Interactive Messaging，安全低延迟交互消息协议）**。Slim 引入了发布-订阅（Pub/Sub）架构并内置安全机制，显著降低了大规模智能体间（Agent-to-Agent）以及智能体与工具间（Agent-to-Tool）的通信时延与开销。 #### 4. 可观测性（Observability） 跨越不同运行平台，生成统一的多智能体追踪数据（Traces）与度量指标（Metrics），全面监控协作流程的运行状态。 **与其他开源生态的兼容** AGNTCY 并未孤立存在，而是与行业现有的开源项目（如 Anthropic 的 MCP 和 Google 的 A2A）深度融合。例如：Agent Directory 支持托管 MCP 和 A2A 记录；Slim 协议支持传输 MCP 和 A2A 消息，确保了行业标准的高度互操作性。 --- ### [15:43](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=943s) 架构阶段与演示案例（Coffee Farm Demo） 整体架构涵盖了智能体生命周期的四个关键阶段：**发现（Discover）**、**组合（Compose）**、**部署（Deploy）**和**评估（Evaluate）**。 ``` +-------------------------------------------------------+ | 1. 发现 (Discover): Agent Directory / OASF Schema | +-------------------------------------------------------+ | 2. 组合 (Compose): A2A / MCP (语法与端点绑定) | +-------------------------------------------------------+ | 3. 部署 (Deploy): Slim 协议 (gRPC Pub/Sub / 低延迟) | +-------------------------------------------------------+ | 4. 评估 (Evaluate): 全局 Trace 与多智能体 Observability| +-------------------------------------------------------+ ``` 在演讲结尾，Tushar 展示了在 Cisco 展台演示的实际应用案例——**咖啡农场（Coffee Farm）多智能体应用**： * **业务场景**：模拟买家向咖啡农场采购咖啡的过程。 * **技术实现**：底层智能体基于 **LangGraph** 构建，智能体之间采用 **A2A** 协议进行沟通，并借助 **MCP** 工具向外部接口查询天气信息。 * **演示重点**： * **Slim 协议的通信性能**：展示了 A2A 消息在 Slim 协议下实现的多对多（Many-to-Many）订阅及一对一（Request-Reply）高效传输。 * **身份验证**：动态校验智能体在操作过程中的执行权限。 * **端到端可观测性**：实时展示请求在多个智能体之间的传递链路与完整的链路追踪（Traces）。 --- - **主讲人**：Tushar Agrawal（Cisco Outshift GenAI 产品管理高级总监） - **开源项目**：AGNTCY（Linux Foundation 托管项目） - **联合发起方**：Cisco Outshift, Dell, Google, Oracle, Red Hat - **涉及核心协议与标准**：OASF, Slim (基于 gRPC 的 Pub/Sub 扩展), MCP, A2A - **核心比喻**：儿童积木协作搭房子（引自 Perplexity） ### [19:02](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1142s) 咖啡供应链多 Agent 协作演示：发布订阅模式 演讲者展示了一段演示视频，用具体的咖啡供应链场景来解释智能体网络（Internet of Agents）的工作机制。在该演示中包含四个主要 Agent：一个买家 Agent（Buyer Agent），以及分别代表巴西（Brazil）、哥伦比亚（Colombia）和越南（Vietnam）三个咖啡供应国的子 Agent。智能体之间通过 A2A 协议进行对话，而底层则基于 SLIM（Secure Local Interaction Mesh）这一通用通信层实现消息互通。此外，哥伦比亚 Agent 还连接了一个 MCP 协议服务器（MCP Server），使其能够接入外部气象应用，从而根据实时天气预测来评估自身的供货能力。 ``` +-------------+ | Buyer Agent | +------+------+ | SLIM (Mesh Layer) +-----+---------------------+ | | v v +------------------+ +-------------------+ | Colombia Agent | | Brazil Agent | | (MCP -> Weather) | | (Info Only) | +------------------+ +-------------------+ ``` 演示的第一个场景展示了“多对多发布订阅模式”（Many-to-Many Publisher-Subscriber Pattern）。买家 Agent 发起广播提问：“当前全球最新的咖啡库存是多少？”虽然此时巴西 Agent 仅获得了提供信息的权限、尚未通过执行具体交易动作的身份验证，但 SLIM 消息层能够确保所有订阅了该主题的 Agent 都能同时收到消息。随后，哥伦比亚 Agent 调用 MCP 服务器获取气象数据，各个供应国 Agent 协同汇总信息，最终将统一的库存报告反馈给买家 Agent。 - **演示角色构成**：1 个买家 Agent，3 个代表供应国（巴西、哥伦比亚、越南）的子 Agent。 - **通信与工具集成**：基于 A2A 进行 Agent 间对话，通过 SLIM 统一消息层传输，利用 MCP 连接哥伦比亚 Agent 与外部 Weather App。 - **消息模式**：通过 SLIM 实现广播式的多对多发布/订阅机制，确保各订阅节点同步接收请求。 ### [21:06](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1266s) 基于身份验证与工具级访问控制的请求响应模式 接下来，演讲者展示了 SLIM 上运行的“请求-响应模式”（Request-Response Pattern）以及基于身份的细粒度安全访问控制（Tool-based Access Control）。在第二段演示中，买家 Agent 发出一个具体的下单请求：“能否以每磅 5 美元的价格从哥伦比亚订购 500 磅咖啡？”买家 Agent 明确指定了目标对象为哥伦比亚 Agent，由于哥伦比亚 Agent 的身份已通过验证，具备处理订单交易的授权，系统顺利为其创建了订单。 随后，买家 Agent 尝试对巴西 Agent 执行相同的订购操作。由于巴西 Agent 的身份仅获准回答咨询信息，未通过执行“下单功能”（Order Function）的身份验证，系统触发了工具级访问控制，直接拒绝了该请求并提示：“巴西 Agent 未通过验证，无法协助处理此特定请求。”这直观地证明了跨 Agent 协作网络中身份验证与权限隔离的重要性。 - **请求内容**：以 $5/磅 的价格向哥伦比亚订购 500 磅咖啡。 - **鉴权对比**：哥伦比亚 Agent 具备订单处理权限，成功下单；巴西 Agent 因未通过交易身份验证被拦截。 - **安全机制**：展示了基于 Agent 身份的工具级访问控制（Tool-based Access Control），防止未经授权的 Agent 执行高风险操作。 ### [22:15](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1335s) AGNTCY 开源倡议与工作组邀请 在展示完演示视频后，演讲者简要总结并呼吁现场观众共同参与 AGNTCY 开源项目的建设。演讲者强调 AGNTCY 是一个开源倡议，秉持中立原则，希望有更多开发者和企业加入（"the more the merrier"），以代表行业内各方的共同利益。演讲者展示了包含多个 QR 码的幻灯片，方便与会者扫码了解详情并加入相关技术工作组。 演讲者进一步介绍了当前设立的各个专项工作组（Working Groups），涵盖身份识别（Identity）、通信网格（SLIM）、可观测性（Observe）以及目录服务（Directory）等核心主题。最后，演讲者邀请大家前往思科（Cisco）展台体验现场 Demo，并进行更深入的技术交流，并在全场的感谢声中结束了本次演讲。 - **开源倡议**：AGNTCY 旨在以中立方式推动 Agent 互联标准，鼓励社区共同建设。 - **专项工作组 (Working Groups)**：包括 Identity（身份）、SLIM（通信）、Observe（可观测性）、Directory（目录服务）。 - **后续互动**：现场参会者可前往 Cisco 展台体验完整 Demo 并与团队展开深度技术讨论。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=4s) 确定性计算向非确定性 AI 智能体的跨越 — Tushar Agrawal > [00:04](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=4s) 传统软件协作依赖于确定的、硬编码的格式；而 AI 智能体拥有自主性（Agency），要求它们能够根据上下文实时理解、反应并动态调整与其它智能体的协作方式。 — Tushar Agrawal > [03:14](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=194s) 想象四个孩子正在一起用积木搭房子。如果一个孩子在某个位置放了一块积木，另一个孩子就必须对此作出反应，并在另一个位置放置相应的积木。 — Tushar Agrawal > [03:14](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=194s) 传统 API 主要是固定的“请求-响应”（Request-Response）模式；而智能体之间的沟通则更加“交谈化”（Chatty），需要具备打断（Interrupt）、纠偏（Course Correct）和进一步澄清（Clarify）的能力。 — Tushar Agrawal > [07:46](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=466s) 如果说互联网时代关注的是虚拟机（VMs），云计算时代关注的是微服务（Microservices），那么智能体时代则需要在应用层之上建立新的语法层（Syntactic Layer）和语义层（Semantic Layer）。 — Tushar Agrawal > [21:06](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1266s) 巴西 Agent 未通过验证，无法协助处理此特定请求。 — Tushar Agrawal > [22:15](https://www.youtube.com/watch?v=HyP9owAr5IQ&t=1335s) the more the merrier — Tushar Agrawal ## Mentioned Tushar Agrawal, Cisco, Outshift, AWS, Perplexity, OSI Model, AGNTCY, Linux Foundation, Dell, Google, Oracle, Red Hat, OASF, MCP, A2A, gRPC, Slim, LangGraph ## Video description Tushar Agrawal, Sr. Director, Product Management - GenAI, Outshift by Cisco Abstracts AI is rapidly transforming software development, and the biggest shift is the move from single model assistants to collaborative multi-agent software. This software brings together agents with different domain expertise, trained on different data sets to solve complex business and scientific problems. Imagine the problems that can be solved when AI agents can discover and trust each other, collaborate seamlessly, and be observed, regardless of how they were built or where they run - this is the promise of the Internet of Agents. Join Tushar Agrawal, Senior Director of Product Management at Outshift by Cisco, to learn about the foundational technologies and standards that the AGNTCY collective is building for the Internet of Agents. In this session you will learn why multi-agent software requires a new OSI model, explore the Internet of Agents architectural components and how they work together with MCP and A2A, and discuss real-world examples of multi-agent workflows. Whether you’re already deploying multi-agent software or just getting started, this session will provide you the insights you need to accelerate your roadmap in this exciting new era of the Internet of Agents. [Source](https://www.youtube.com/watch?v=HyP9owAr5IQ)

[0:04] Tushar Agrawal: 确定性计算向非确定性 AI 智能体的跨越 [0:04] Tushar Agrawal: 传统软件协作依赖于确定的、硬编码的格式；而 AI 智能体拥有自主性（Agency），要求它们能够根据上下文实时理解、反应并动态调整与其它智能体的协作方式。 [3:14] Tushar Agrawal: 想象四个孩子正在一起用积木搭房子。如果一个孩子在某个位置放了一块积木，另一个孩子就必须对此作出反应，并在另一个位置放置相应的积木。 [3:14] Tushar Agrawal: 传统 API 主要是固定的“请求-响应”（Request-Response）模式；而智能体之间的沟通则更加“交谈化”（Chatty），需要具备打断（Interrupt）、纠偏（Course Correct）和进一步澄清（Clarify）的能力。 [7:46] Tushar Agrawal: 如果说互联网时代关注的是虚拟机（VMs），云计算时代关注的是微服务（Microservices），那么智能体时代则需要在应用层之上建立新的语法层（Syntactic Layer）和语义层（Semantic Layer）。 [21:06] Tushar Agrawal: 巴西 Agent 未通过验证，无法协助处理此特定请求。 [22:15] Tushar Agrawal: the more the merrier

[0:00] Tushar Agrawal: Imagine the problems that can be solved when AI agents can discover and trust each other, collaborate seamlessly, and be observed, regardless of how they were built or where they run - this is the promise of the Internet of Agents. [0:04] Tushar Agrawal: Outshift functions as Cisco’s emerging technology incubation group, exploring adjacent domain innovations that eventually feed into existing or new Cisco product portfolios to drive customer success. [1:57] Tushar Agrawal: Transitioning from traditional cloud computing to agentic software requires a fundamental mental shift. [1:57] Tushar Agrawal: Unlocking the full promise of AI agents requires allowing them to collaborate autonomously across these heterogeneous environments without relying on hardcoded, brittle API integrations. [4:31] Tushar Agrawal: Standard REST APIs lack this fluid interactivity. Agents require a conversational, low-latency interface capable of handling interrupts, clarifications, and iterative multi-party dialogues. [4:31] Tushar Agrawal: To bridge the gap between software-scale execution and human-like reasoning, Cisco advocates for a new global architecture: the Internet of Agents. [9:38] Tushar Agrawal: AGNTCY aims to construct vendor-neutral protocols that prevent lock-in across the agent ecosystem. [19:02] Tushar Agrawal: What is the latest global coffee inventory? [21:05] Tushar Agrawal: Can I order 500 pounds of coffee from Colombia at $5 a pound? [22:15] Tushar Agrawal: the more the merrier

[0:04] So good to see you all here. Thanks for [0:06] uh sparing your time for this session. [0:08] Um hopefully you'll walk out with um a [0:12] good sense of some of the challenges [0:14] that we see as important as agentic [0:18] collaboration becomes more and more [0:20] prominent and uh some of the work that [0:23] uh we're working with a whole consortium [0:25] of industry [0:27] um to to address that. So hopefully [0:29] they'll be educational from that lens. [0:31] Um just to introduce myself uh I'm [0:33] Tushar. I uh lead product management for [0:37] outshift [0:38] uh which I'll talk about in a second. [0:40] It's one of the uh incubation groups [0:42] within Cisco. Um I have about 15 years [0:45] of experience in product management uh [0:47] both across large companies and [0:49] startups. And uh my uh journey into AI [0:55] uh started uh at AWS uh where I led uh [0:59] product management for uh their computer [1:01] vision suite of services. So uh good to [1:05] be here with you all. Um our shift so [1:08] want to kind of quickly talk about what [1:09] our shift is because that will kind of [1:11] give you the context for why we are [1:13] working on this area. Um our shift is [1:15] Cisco's incubation team. uh we work on [1:18] delivering uh what's next uh for Cisco [1:21] in emerging technology areas in adjacent [1:25] spaces that make sense for Cisco [1:27] businesses to get into and u in service [1:30] of that uh we look at vast we look at a [1:33] vast variety of u you know areas and and [1:37] um technology uh innovations which then [1:41] find its way uh to Cisco uh through the [1:44] various uh products be it existing [1:46] portfolios or be it new portfolios that [1:48] we can uh you know set up uh as in [1:51] pursuit of you know customer success and [1:53] so on. So [1:57] um let's start with sort of some of the [1:59] setup here for why um you know um you [2:04] sort of need to think about the entire [2:06] space of uh agent collaboration and um [2:10] agent interoperability a bit differently [2:13] than perhaps we have looked at in terms [2:15] of how you know microservices and VMs [2:19] have collaborated in the past. Um the [2:22] biggest difference is that uh these [2:24] agents are supposed to have agency [2:28] right and in service of agency you can't [2:33] really apply the same deterministic [2:36] formats that were applied to [2:39] deterministic software collaboration. uh [2:41] inherently there is a requirement for [2:44] these agents to be able to understand [2:47] react um and uh uh you know sort of [2:51] modify the way they are approaching a [2:53] collaboration with another agent um [2:56] relative to what we used to see in the [2:58] deterministic world. So we need to sort [3:01] of think about what does this move from [3:04] the cloud computing which is currently [3:06] largely in the deterministic era looks [3:08] like when you get into the [3:09] non-deterministic part. [3:12] So [3:14] um we know that a lot of the tasks that [3:17] are going to be required uh of agents [3:21] and multiple agents are going to need [3:24] agentic collaboration. uh if you [3:26] approach this from a user lens, if I [3:28] need to get a task done today on let's [3:31] say my own work as part of my own work, [3:34] I probably am looking at various [3:36] different tools and applications that I [3:38] need to do that work across and so it's [3:42] hard to uh hardcore it, hard to make it [3:45] deterministic if you want to unleash the [3:46] full promise of agents. And so we think [3:50] that uh agents will absolutely need to [3:52] collaborate because a lot of tasks that [3:54] will be given to them would be um you [3:56] know would require different patterns of [3:57] collaboration. And we know that these [4:01] agents like the example I gave will come [4:04] from various different vendors. They [4:07] will use different development [4:09] frameworks. They'll be built on [4:11] different clouds. Um they'll have [4:14] different identity providers. [4:17] They'll belong to different [4:18] organizations [4:19] and they'll have different profiles and [4:21] personas which will all have to be uh [4:23] dynamically adjusted depending on the [4:26] task at hand. Right? And so um we think [4:31] that there is a need for a new internet [4:35] and we like to call it the internet of [4:37] agents. [4:38] um where we have the ability to blend [4:43] some of the software-l like capabilities [4:45] of agents where agents are fast, they [4:48] can scale massively uh but also some of [4:51] the human-like attributes of agents [4:53] which are in the sense of um uh you know [4:57] the dynamism, the agency, the ability to [4:59] reason uh the ability to uh understand [5:03] uh what the task at hand is etc. And we [5:06] think that this internet of agents is um [5:13] where Cisco as such and you know all the [5:15] industry players can play a huge role [5:18] just like Cisco played a huge role in [5:20] standardizing the stack for the internet [5:23] of the you know web era right which is [5:26] you you create sort of a series of [5:27] standards and protocols that enable [5:29] these systems to understand collaborate [5:33] uh discover and identify Okay. So, just [5:36] like DNS works today, you have a URL, [5:39] you have a and and sort of is able to [5:41] map it to wherever that web is hosted. [5:43] How do you build the ability for an [5:45] agent to be able to discover other [5:47] agents, trust them, identify them, make [5:51] sure that they're doing their work in in [5:54] adherence to the policies that are [5:55] required in that moment? How do you make [5:57] sure that these agents can talk to each [5:59] other? And this is where you know I was [6:01] actually uh uh doing a search on [6:04] perplexity on to how to explain that in [6:06] an easy way and u you know the example [6:11] that perplexity gave is let's say there [6:13] are four kids that are collaborating on [6:15] building a house right now if these kids [6:18] are collaborating on building a house uh [6:21] it's very much like what these agents [6:24] would need to collaborate on so you need [6:25] to understand what the house looks like [6:28] to understand um uh you know where if if [6:31] if one kid puts a block in one place the [6:34] other kid has to react to that and put [6:35] the block in another place and so on. So [6:38] APIs for example traditionally don't [6:40] have that level of flexibility right [6:42] there is more of a um you know request [6:45] response pattern versus agents require [6:48] the ability for other agent to interrupt [6:51] to course correct to uh clarify and it's [6:54] a lot more chatty interface and and [6:56] that's an example of you know some of [6:58] the new uh aspects which you know A2A is [7:02] looking at for example we had a protocol [7:04] called agent connect protocol uh so [7:06] there's a need for innovations that are [7:09] required on how can you enable these [7:11] agents to communicate with each other [7:13] that are over and above what an API can [7:15] offer today. And then lastly, it's a [7:18] it's a loop. You you have to be able to [7:20] observe and evaluate how things are [7:22] going and then react, right? So there's [7:25] also a need in this multi-agentic, [7:27] multi- vendor world, multi-environment [7:29] world for these agents to be able to [7:31] collaborate with each other to observe [7:34] it in a in a in a sort of unified [7:36] holistic fashion and um figure out how [7:39] what the next uh iteration needs to look [7:42] like. So to address uh this you know we [7:46] sort of looked at how should the OSI [7:49] model be enhanced right and so I think [7:51] all of you are sort of familiar with the [7:53] OSI model. has served us very well so [7:55] far. Uh you've got the physical data [7:58] network and transport layers at the very [8:00] bottom. You've got the uh session [8:02] presentation and application layers at [8:04] the top. U but we think that there is [8:07] now a need for syntactic and semantic [8:11] um you know alignment that can enable [8:13] this agentto agent collaboration. And so [8:16] we had uh you know the internet era that [8:18] was more focused on VMs. We had the [8:21] cloud era that was sort of more focused [8:22] on microservices. I think we we think [8:25] that we need an internet of agents the [8:27] next layer on top of that that deals [8:30] with some of the syntactical things like [8:32] MCP like A2A but then also enables the [8:35] path for some of the semantic aspects [8:37] which is you know how do you for example [8:39] take a user request and break it down [8:42] into uh what permissions will this [8:45] request require and make it a part of [8:46] the plan uh not an afterthought. Right? [8:50] How do you make a plan for your agentic [8:52] collaboration in the sense that it [8:53] incorporates some of the semantic [8:54] understanding of the task itself? [8:57] And that's what uh we've been focused [8:59] on. Uh so we uh you announced this [9:02] initiative earlier this year. um it's [9:06] called the internet of agents and uh we [9:09] are focused on making sure that this [9:12] internet of agents is an open [9:14] interoperable internet for agentto agent [9:17] collaboration across you know wherever [9:20] these agents are getting built right so [9:25] um so let's get into some of the details [9:27] on what this initiative is all about and [9:29] then we'll get into some technical [9:30] details on u you know what are some of [9:33] the components and pieces that we've [9:35] been working on so far. [9:38] So, uh this initiative was called [9:40] agency. Uh and it's not a misspelling. [9:43] This is what it's actually called. The e [9:45] is deliberately missing. Um it's uh now [9:48] a Linux foundation project. Um it's an [9:52] open source project where over 75 [9:55] startups have participating and are [9:57] participating in enabling this multi- [10:00] vendor multi-agentic collaboration you [10:03] know framework and standards uh that are [10:06] required across you know various uh [10:09] suits of capability. Um the formative me [10:12] members who joined us in this initiative [10:14] are u obviously we played a role in [10:17] outshift in kind of driving this but we [10:19] also had uh team members from Dell, [10:22] Google, Oracle and Red Hat who have [10:24] joined uh this effort with us and we [10:27] invite all of you who are interested in [10:29] contributing to this initiative in the [10:30] open source uh to enable the same kind [10:33] of flexibility and interoperability that [10:35] we have with the internet today in the [10:36] agentic world to come and join this and [10:38] we'll share some of the details on how [10:40] you can do that. subsequently. [10:43] So, let me do a quick uh question and [10:46] answer here just a show of hands. How [10:49] many of you are actually working on [10:51] multi-identic apps uh at this point? Any [10:55] any hands raised? [10:58] Okay, cool. And anyone who is currently [11:01] looking at uh multi-agentic apps that do [11:04] require cross vendor multi-agentic [11:07] collaboration. [11:09] Okay, great. So, some of you so I think [11:11] uh that's great because uh you know we [11:13] are then working on this at the right [11:15] time and so hopefully some of these [11:17] standards can help that uh journey uh [11:19] you know make that journey easy for you [11:20] as we go along. [11:23] So let's look at what are some of the [11:25] components that we are working on right [11:28] so there are four main aspects and this [11:30] kind of flows from some of the things [11:31] that we talked about earlier [11:34] um as to what are some of the challenges [11:36] that you can encounter when you have [11:39] multiple multiple agents that are from [11:42] different vendors trying to collaborate [11:43] with each other. The first is agent [11:46] directory. [11:47] Right? So when you have a task that is [11:50] given to an agent or a composer of an [11:52] agent or an orchestration uh agent, uh [11:55] the first thing that it would need to do [11:56] in due time when you have these tasks [11:59] that are not sort of predetermined would [12:01] be to go find other agents and [12:04] understand the capabilities of those [12:05] agents. And so you need the ability to [12:08] describe an agent in a way that any [12:11] other agent can read and understand, [12:13] right? And so today what you have is you [12:15] know all of these agents are being [12:17] described in their own vendor specific [12:19] manner. How do you really get to a [12:22] standard you know framework uh which we [12:24] have launched called OASF which is open [12:27] agentic schema framework. How do you [12:29] describe an agent in a consistent way so [12:32] that uh different agents can discover [12:35] them can learn about their capabilities [12:37] can learn about how to invoke them can [12:40] learn about how to configure them and so [12:42] on. So, agent directory takes OASF as [12:46] the base and it enables you to set up a [12:49] decentralized agent registry model where [12:54] agents that are in your enterprise [12:56] organization and hosted by other [13:00] organizations be it a vendor or be it [13:02] other enterprises have a way to discover [13:05] each other in a decentralized way to [13:06] understand each other and collaborate. [13:09] Right? So uh so that's the focus of [13:11] agent directory and we'll talk about [13:12] that in a bit more if I have time. Um [13:15] agent identity. So the same issue comes [13:18] up when you have multi-agent [13:19] collaboration. How do you so you know [13:22] you have one agent that is using a [13:24] particular IDP which is an ID provider. [13:26] You have the other agent that is using [13:27] another identity provider. Uh there's [13:30] one agent that is allowed to do is [13:32] capable of doing 10 tasks. The other [13:34] agent is capable of doing five tasks. [13:36] when they engage with each other, what [13:38] tasks should you allow them to do, [13:41] right? What tasks how do you make sure [13:44] that these different IDPs are still uh [13:46] you know able to sort of you able to [13:48] sort of sort through that and get these [13:50] things to interact. So that's what [13:51] identity focuses on. Messaging is the [13:54] fourth component that we are actively [13:57] focusing on which is more focused on [14:00] when you have these agents ensemble. So [14:02] you have five agents, 10 agents. How do [14:05] you enable the collabor the the [14:07] messaging of this agent to happen in a [14:09] lot more secure and efficient way? And [14:12] here we are extending uh gRPC [14:15] uh to make it more publisher subscriber [14:18] oriented uh while also bringing some [14:20] elements of inbuilt security so that uh [14:23] the overhead that is currently [14:24] associated at large scale multi-agentic [14:27] communication both for let's say agent [14:29] to agent or for agent to tool uh can [14:33] also be addressed and we'll talk about [14:35] that. We call it slim which is a secure [14:38] low latency interactive messaging [14:40] protocol and and that's what we're [14:42] working on here. And then finally [14:44] observability right which is uh how do [14:46] you make sure that you can generate [14:48] traces regardless of which platform the [14:50] agents are running on that you can [14:52] generate multi-agentic metrics that tell [14:56] you how uh this agentic multi-agentic [14:59] operation is running. [15:01] And then lastly, we are doing this um in [15:04] conjunction and in collaboration with [15:06] other open source initiatives such as [15:08] MCP and A2A. And so you will see as we [15:12] go along in this, you know, a little bit [15:14] more technical that directory, for [15:17] example, is able to host MCP records. [15:20] It's able to host A2A records. Uh Slim [15:23] is able to have A2A messages go over [15:27] Slim and MCP messages over Slim. And so [15:30] we're trying to sort of work with the [15:32] industry here to make sure that we're [15:34] defining standards that are [15:35] interoperable not just between the [15:37] agency project or within the agency [15:39] project but also across multiple [15:40] projects that are out there. [15:43] Um I want to check how am I doing on [15:45] time. Okay, cool. So um this is what the [15:51] um architecture looks like at a high [15:53] level, right? Um so you have [15:57] the four phases that we are really sort [15:59] of focused on right now. The discover [16:01] phase which is how do agents discover [16:03] each other. Uh the compose phase um how [16:07] do you then bring them together and get [16:09] them to uh do a task. Um the u deploy [16:13] phase and the evaluate phase. And so if [16:16] you look from the top to the down you'll [16:18] see that the agent directory is really [16:20] focusing on the discovery phase. And um [16:23] you know you can think of agent [16:25] directory as a a decentralized service [16:27] where these agent records are um you [16:31] know available uh in the standardized [16:33] format which you call OASF and uh what [16:38] agent directory allows then is a [16:40] decentralized syncing of this but also a [16:43] secure um uh you know ability to sort of [16:46] have lineage in terms of when these [16:49] records were created and who is the you [16:51] know the the rightful owner of that [16:53] record etc. [16:55] Then you come down into the diagram you [16:57] see the agents [16:59] you know that are collaborating and we [17:01] talked about the semantic and the [17:02] syntactic layer there you got A2A and [17:05] MCP that are sort of more focused on the [17:07] syntactic part of it which is you know [17:09] how do you make sure that the agents [17:12] have the right endpoints that can be [17:14] used to invoke each other but then also [17:16] um the slim layer at the bottom which [17:20] enables these MCP and A2A messages to be [17:23] sent uh to each other in a highly u you [17:26] know efficient secure low latency [17:28] fashion. [17:30] So um with that I think uh I can [17:33] probably skip to um I did have some more [17:36] slides on the details but let me kind of [17:37] skip through a quick demo here. Uh we do [17:40] have a a booth at the Cisco uh booth [17:44] area. Um you can see this demo at the [17:48] booth. Uh it's a multi-agentic [17:50] application which simulates a coffee [17:53] farm [17:55] um and a buyer wants to come and buy [17:57] coffee and there are multiple agents uh [18:00] that are involved in that and and those [18:02] agents are currently built using [18:04] langraph uh you know they collaborate [18:06] with each other using A2A as on the [18:08] agentto agent uh interaction and they [18:10] collaborate with an MCP tool uh to check [18:13] the weather for example. [18:15] Um so what we show in this demo is the [18:18] ability for uh these agents to [18:19] communicate using A2A over slim um both [18:24] in a many to many format um but also in [18:27] a request reply onetoone pattern and so [18:30] you can kind of see how slim enables you [18:33] know this many to many and onetoone [18:35] collaboration. We also show agent [18:37] identity and how we are using that to [18:39] verify the permissions of what an agent [18:42] is supposed to do or is able to do. And [18:45] then um we kind of have the u observe uh [18:50] portion integrated so that you can see [18:52] uh you know what the traces look like. [18:54] How did the request uh request response [18:57] u you know emerge for this particular [18:58] application. So [19:02] let's see. So let me go into the demo [19:04] here and uh show that to you. So [19:07] hopefully uh this can play okay. [19:12] Um so what's really happening here? you [19:13] can read this on the top on the top [19:15] left. Um so the first question that you [19:18] know the the uh so there are four as I [19:20] said there are sort of four agents so [19:22] there the buyer agent um and there is [19:24] there are sort of three uh countries uh [19:27] in such a Brazil colombia and Vietnam [19:30] who are able to supply uh coffee and uh [19:35] there's an interaction happening between [19:36] these uh the buyer agent and these [19:39] multiple uh you know sub agents uh to [19:42] serve the request that the user has and [19:44] what you see here is that this [19:45] interaction action is happening. Uh you [19:47] know A2A is is how these agents are [19:49] talking to each other and slim is the uh [19:52] common layer over which this [19:53] communication is happening. Uh there's [19:55] also an MCP server that is connecting [19:57] the uh Colombia agent to a weather app [20:00] right so that it can use that weather [20:02] information to uh provide information on [20:05] you know whether I'm able to supply some [20:06] quantity or not. And so the first [20:09] question that uh this buyer agent asks [20:11] is u you know what is the latest uh uh [20:14] global coffee inventory. And in this [20:16] case you'll see that even though the uh [20:19] Brazil agent is not sort of validated [20:22] per se uh it's not validated to perform [20:24] an action where it's validated right now [20:25] to perform uh you know provide [20:27] information. And so in this case if I [20:30] play it forward [20:32] you'll see that when the request goes [20:34] through it it slim it slim makes sure [20:36] that all of these agents who have [20:38] subscribed to this particular uh you [20:40] know topic were able to get the message [20:43] at the same time. [20:45] Um and then the col uh the Colombia [20:48] agent also used the MCP server and then [20:51] they all sort of got together and [20:52] responded the response back to the [20:54] buyer. [20:55] Right? So that's the u that's sort of [20:57] one example of um how you're doing sort [21:01] of a many to many publisher subscriber [21:03] pattern. Um now you'll see how you can [21:06] do a request response pattern over slim. [21:09] And so let's say the question here is [21:11] can I order uh 500 pounds of coffee from [21:15] Colombia at $5 a pound? Right? And so in [21:17] this case, it specifically called out [21:20] Colombia um um as the agent that it [21:23] wants to use and and Colombia was [21:25] validated in terms of its identity to [21:27] provide this uh response. And so it goes [21:30] ahead and uh creates an order for it. [21:41] Right? And now you'll try doing the same [21:43] exercise with Brazil. uh in this case [21:46] because Brazil has not been validated [21:49] for the identity for sort of performing [21:51] an order function uh we are able to do [21:53] sort of like a toolbased access control [21:56] here to say uh no right like hey Brazil [22:00] has not been validated and um Brazil [22:03] cannot help you with this specific [22:05] request right so [22:08] um so that's why I wanted to show you u [22:10] with that I think we are up on time so [22:13] if there are any question oh actually I [22:15] I do want to show you one more slide. So [22:17] um uh you can sort of scan some of these [22:20] QR codes to learn more about agency. Uh [22:23] we really invite you to come join us. Uh [22:25] this is an open-source initiative. Um [22:28] the more the merrier the better in terms [22:30] of representing all your interests uh [22:32] making sure that it's done in a neutral [22:35] uh way and uh we also invite you to join [22:38] some of the working groups uh that we [22:40] have set up on these topics. So if [22:41] you're interested in identity or slim or [22:44] observe or uh directory uh you can come [22:47] in and start joining these working group [22:49] meetings and participate in that. Also [22:52] please check out the Cisco booth area to [22:54] see this demo and have a more detailed [22:56] technical conversation on that. So with [22:58] that thanks a lot for your time. [23:00] Hopefully it was educational and I'm I'm [23:02] around in the corner here if you have [23:03] any questions for