Cox, Hyland Executives Warn Token Costs Threaten Enterprise AI Scale

Cox与Hyland高管警告：Token成本威胁企业AI规模化 除非企业解决数据碎片化问题并优先针对昂贵的外部合同，否则自主企业智能体的运营成本可能超过其创造的价值。 与会嘉宾抛开行业炒作，审视了部署智能体系统的现实情况，强调未加管制的Token支出和复杂的工具编排极易导致智能体的运营成本高于其创造的价值。实现有效规模化需要稳固的非结构化数据架构、统一的全局命名空间以及可组合的层级结构，而非孤立的“母体智能体”或无序蔓延的单点解决方案。讨论指出，企业成功落地的关键在于规范的沙箱机制、本地与边缘混合架构，以及侧重于自动化外部承包商任务和赋能员工而非裁减人员的组织变革管理。 单体经济效益与 Token 治理是企业扩展智能体系统的主要瓶颈，失控的自主推理和多步骤工具调用极易导致运行成本超出其创造的业务价值。 (4:00) 企业智能体的核心痛点并非基础模型推理，而在于跨传统遗留系统对高度分散且庞大的非结构化数据进行受控的联邦式编排访问。 (6:12) 随着计算机间自主通信架构普及，建立统一的全局数据命名空间成为打破算力、电力和存储壁垒、规避数据孤岛的关键基础设施要求。 (8:26) 企业应当摒弃片面追求智能体数量的虚荣指标，转向构建模块化、层次化的可组合本体架构，由中央编排智能体统一协调各细分功能子智能体。 (10:41) 在商业化落地与投资回报方面，企业优先通过构建概念验证（POC）智能体取代外部外包合同与高成本第三方供应商，以实现立竿见影的降本增效并将资金反哺于自主研发。 (28:14)

Enterprise technology leaders across automotive, telecommunications, content management, and storage infrastructure argue that scaling autonomous agents requires prioritizing foundational data access, unit economics, and hybrid infrastructure over raw model hype. Rather than pursuing unconstrained agentic meshes, organizations are adopting composable orchestrations, sandboxed internal proving grounds, and disciplined data governance to secure predictable returns. Token Economics and Pragmatic System Architecture (4:00) Panelists from Cox Automotive and Cox Communications emphasize that enterprise deployment of autonomous agentic systems remains constrained by operational economics and token governance. Jodi Blomberg highlights that without strict scaffolding, unpredictable multi-step reasoning can easily result in agentic systems costing more to operate than the value they generate. Organizations must therefore absorb higher token expenditures during exploratory phases to identify repeatable patterns before optimizing costs for production. Matthew Shorts cautions against vanity metrics, noting that claims of deploying 100 enterprise agents typically decompose to roughly ten core task agents supported by minor utility sub-routines. To mitigate commercial and technical risk, organizations prioritize internal developer and operational workflows as proving grounds before deploying customer-facing tools. Blomberg and Shorts observe that validating agents internally enables teams to absorb architectural failure modes—such as hybrid cloud networking and data egress overruns—without breaching external customer service level agreements. Unstructured Data Governance and Global Access (6:12) Enterprise content management and storage specialists identify data orchestration as the primary bottleneck to autonomous agent execution. Tim McIntire of Hyland Software notes that approximately 80% of enterprise information resides in unstructured formats across disparate legacy systems, making federated data access a more urgent challenge than raw model inference. With global data generation projected to reach 300 zettabytes by 2027, panelists observe that data governance has entered a critical phase of operational hygiene to prevent unmanaged repositories from becoming liabilities. To overcome data gravity and security silos across on-premises, edge, and cloud environments, Sam Newnam of Hammerspace advocates establishing a unified global namespace. A consolidated data access layer allows organizations to enforce consistent permissions and metadata indexing across distributed environments. Newnam also stresses the need for end-to-end observability—using tools like Zipkin tracing and mdtest metadata benchmarks—to identify latency bottlenecks in retrieval-augmented generation pipelines and inform cost-effective data placement decisions. Hybrid Deployment, Edge SLMs, and Protocol Standards (31:49) Rather than relying exclusively on monolithic cloud language models, panelists point to an architectural transition toward small language models running locally on edge hardware for low-latency, deterministic tasks. Emerging hybrid architectures pair turnkey on-premises stacks—such as Dell AI Factory with NVIDIA NIM microservices—directly with private data warehouses, while selectively synchronizing vector embeddings to the cloud for compute-heavy workloads. Panelists note that the enterprise management stack and GPU allocation logistics, rather than underlying base models, dictate price-performance. Scaling also demands standardized interoperability to replace proprietary vendor silos. The panel highlights initiatives such as the Model Context Protocol and Google's A2A frameworks as mechanisms for server-to-server tool execution. However, speakers caution that industry-wide governance frameworks for agent communication remain unratified, raising security concerns regarding whether autonomous GPU-to-GPU pathways could inadvertently expose sensitive corporate data. Governance, Contract Displacement, and Workforce Augmentation (28:14) Addressing organizational impact, the panel argues that enterprise agent scaling centers on augmenting workers rather than eliminating internal headcount. In more than 90% of production workflows, human-in-the-loop oversight remains essential to manage edge cases and eliminate routine process errors. To capture rapid returns without internal friction, enterprises target external spend: one AI accelerator identified approximately $20 million in operational efficiencies across 27 proofs of concept within six months by replacing third-party agency retainers and offshore data-processing contracts. For continuous management, panelists propose capability taxonomies and supervisory ethics agents to evaluate subordinate task agents against corporate standards. Sandboxed environments with strict copy data management, fine-grained access controls, and red-team security assessments are treated as mandatory prerequisites before granting autonomous agents broad execution authority across core systems. While enterprise interest in autonomous agents is widespread, sustainable scale depends on resolving unit economics, unratified communication protocols, and federated unstructured data governance. Until unified interoperability and real-time security guardrails mature, organizations remain reliant on tightly bounded hybrid architectures, internal proving grounds, and supervisory human oversight.

企业在规模化落地自主智能体（Agent）系统时，正从概念炒作转向对单位经济学与底层基础设施的务实审视；专家指出，规模化瓶颈不在于基础模型推理，而在于非结构化数据访问治理、网络与算力重力平衡，以及通过替代外部采购合同而非内部裁员来验证真实的投资回报率。 规模化现实与单位经济学考量 (4:00) Cox Automotive 人工智能与机器学习副总裁 Jodi Blomberg 认为，当前业界关于大规模跨部门“智能体网络”已成熟落地的说法言过其实。由于自主工具执行和多步推理过程具备高度不确定性，企业极易构建出一个运行成本远超产出收益的智能体系统。Blomberg 强调，当前行业仍处于动态摸索阶段，企业应采取务实策略：在前期探索阶段容忍较高的 Token 消耗以验证可行性，随后必须运用传统机器学习的严谨工程标准，将经验提炼为确定性的复用架构，并实施严格的成本优化与 Token 管控。 Cox Communications 的 Matthew Shorts 亦对虚荣指标提出预警。他指出，许多组织宣称部署了上百个智能体，实质上往往只是约 10 个核心任务智能体配合数十个小型实用子程序。为避免在碎片化的厂商市场中被套牢，企业不应押注单一的通用“母体智能体”，而应建立清晰的分层本体架构，通过中央协调型智能体管理具备明确职责与边界的微型子智能体，在小规模、低风险业务场景中逐步沉淀可扩展的标准单元。 破解非结构化数据重力与访问瓶颈 (8:26) Hyland Software 首席技术官 Tim McIntire 表示，企业中约 80% 的信息属于文档和富媒体等非结构化数据。企业级智能体面临的关键卡点并非算法调用，而是如何在海量且分散的遗留记录系统中，建立起受控、合规的联邦数据访问机制。面对全球数据总量预计在 2027 年达到 300 ZB 的增长态势，智能体以数百万 FLOPs 的速度高速吞吐信息，若企业未能事先梳理杂乱无章的遗留数据孤岛，数据资产将反向演化为巨大的运行风险。 针对算力分散和数据迁移难题，Hammerspace 的 AI 解决方案高级总监 Sam Newnam 提出通过“全局命名空间”（Global Namespace）打通本地文件存储、云端对象存储与边缘环境。该架构向智能体提供跨地理位置与云厂商的一致访问层和安全边界，规避数据重力约束。此外，团队强调端到端遥测的重要性，主张利用 Zipkin 追踪及 mdtest 元数据性能基准等工具排查 RAG 和跨节点协作中的网络延迟，以便在边缘低延迟处理与可扩展云存储之间做出理性的成本与架构权衡。 架构治理、沙箱隔离与监督机制 (19:05) 智能体获得系统操作权限后，机器间的毫秒级交互使企业安全和数据外泄防护面临全新挑战。虽然 Model Context Protocol（MCP）等协议推动了跨服务工具调用，但全行业范围内的权威治理标准尚未确立。为了防范未经授权的数据通路与失控调用，专家组一致主张在生产推广前设置严密的隔离“沙箱玩场”（Playgrounds），针对数据副本复制管理、细粒度权限和红蓝对抗测试建立刚性护栏。 在架构分层方面，Jodi Blomberg 提出引入组织级“共享伦理智能体”或“价值观智能体”构想，将其作为贯穿全局的治理层，持续评估子智能体行为是否符合企业合规原则。企业软件架构必须依照业务领域划分能力分类学（Taxonomy），同时在后台长期部署专门的技术与监管智能体，负责审计、可观测性与访问阻断，落实“具备执行能力并不代表应当执行”的安全底线。 替代外部合同与人机协同组织变革 (28:14) 针对董事会与管理层对劳动力成本削减的关注，专家组指出，AI 加速器的重心主要在于锁定速赢的运营效率，而非直接削减内部员工编制。企业最有效的切入点是通过概念验证（POC）替代昂贵的外部供应商合同，例如第三方面向营销的内容生成机构，或是专门负责 PDF 录入与数据提取的离岸业务流程外包（BPO）团队。以某企业为例，其在半年内通过 27 个 POC 锁定了约 2000 万美元的运营效率收益，并将节省资金重新注回核心技术与产品研发。 在组织协同上，与会者指出超过 90% 的生产流程仍然必须保持人机协同（Human-in-the-loop），智能体的核心价值是剥离重复繁琐的人工差错，释放员工处理高价值创新业务。企业推行 AI 无法仅靠浅层的非正式分享会，而需要深入一线——例如安排技术团队跟访现场汽修车队以摸清底层痛点——并通过在各职能部门嵌入专职推广代表，打破将生成式 AI 视作“高级版 Google”的认知局限，建立自上而下推动与自下而上理解相结合的转型路径。 边缘小模型与混合云部署范式 (31:49) 智能体底层技术栈正从完全依赖中心化云端大型 LLM，向轻量化、分布式的混合架构过渡。面板指出，本地边缘硬件部署的小语言模型（SLM）因响应迅速、成本低廉且具备高确定性，正成为处理端点任务的关键方案。同时，行业正从非标准的临时胶水代码逐步迈向以 MCP 和 Google A2A 为代表的互操作协议规范。 在数据主权与云出口成本（Egress）的双重约束下，企业正转向将敏感核心资产保留在私有数据仓库，并利用如 Dell AI Factory 搭配 NVIDIA NIM 微服务在本地进行计算的模式；非敏感数据则通过向量化嵌入（Embeddings）同步至公共云进行弹性扩展。研讨强调，混合架构中决定性价比与吞吐瓶颈的决定性要素，已不再是底层基础大模型本身，而是企业级的管理软件调度层与 GPU 资源分配逻辑。 当前企业智能体规模化仍面临非结构化数据散乱、端到端可观测性缺失以及行业安全治理标准缺位等现实瓶颈。尽管通过削减外部外包开销为企业指明了清晰的前期投资回报路径，但如何在保持人机协同安全边界的同时，平衡多智能体交互引发的算力成本与网络开销，依然是多云环境下长期部署的未决难题。

## Background 该内容由 YouTube 频道 Ai4 于 2025 年 8 月 24 日发布。视频记录了一场关于企业级 AI Agent 规模化基础设施的圆桌讨论，汇聚了来自 TIAA、Cox Automotive、Cox Communications、Hyland 和 Hammerspace 的多位技术高管，共同探讨支撑智能体在企业中落地所需的数据架构、算力配置、治理与管理挑战。 ## TL;DR 圆桌嘉宾围绕企业部署 AI Agent 的现实挑战展开深入探讨，指出单纯追求智能体数量往往陷入经济账无法跑通和炒作夸大的陷阱。专家们强调，Agent 规模化的核心壁垒不在于模型本身，而在于解决非结构化数据治理、构建跨区域的全局命名空间与联邦数据访问平台，并建立端到端可观测性与权限安全红线。在落地策略上，企业应倡导人机协同并优先解决内部员工与现场高频痛点，通过削减第三方外包服务快速释放 ROI，同时积极拥抱端侧小模型、模型上下文协议（MCP）及混合部署等前沿架构趋势。 ## Atmosphere & style 讨论开场伴随着轻松诙谐的趣事破冰，随后迅速切换为硬核、务实且充满审慎怀疑的技术交锋。嘉宾们毫不避讳地戳破行业营销泡沫，坦率直言“自欺欺人”与“踩坑留下的伤疤”，展现出一线技术管理者在面对巨额 Token 成本与数据混乱时的清醒与韧性。整场对话既有对底层基础设施演进的深度洞察，又洋溢着同行之间相互打趣与求真求实的极客共鸣，为观众呈现了一场褪去浮华、直面落地痛点的实战思辨。 ## Key takeaways - [03:55](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=235s) 企业在智能体规模化部署上面临严峻的经济账挑战，很多大规模落地宣称并不现实，当前应采取‘战略下注’与小步验证策略，在探索出可行闭环后回归传统机器学习的标准化打法控制 Token 成本。 - [06:44](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=404s) 真正的 Agentic 工作流具备自主性、持续运行与免人工介入的计算智能三大特征；构建企业级平台的核心壁垒在于非结构化数据管理与多记录系统的联邦安全访问。 - [08:28](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=508s) 机器对机器（M2M）交互使权限审计与可观测性更为复杂，企业需构建跨云和地理位置的‘全局命名空间’，并在基础设施层部署自主 Agent 优化网络与数据延迟。 - [10:04](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=604s) 避免将细小工具夸大为独立 Agent，应基于本体（Ontology）构建可组合架构，依靠编排智能体（Orchestrator）跨部门调度微型用例与复用组件。 - [12:02](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=722s) 推广智能体需普及全员 AI 素养并设立业务布道者，在给予团队自由探索沙盒的同时建立安全合规治理，平衡创新速度与架构标准化。 - [15:08](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=908s) 落地实践采取三步走策略：率先在自身研发流程与内部现场业务中试错踩坑，将沉淀的‘伤疤’与架构经验反哺并简化外部商业产品的安全交付。 - [19:05](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1145s) 上下文工程正演变为自动化企业引擎，但模型上下文协议（MCP）等工具调用标准在安全方面尚在早期，必须严格把控自主 Agent 的权限边界与数据泄露隐患。 - [21:05](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1265s) 规模化扩展的关键约束是性价比（Price Performance）与投资回报率（ROI），算力虽逐渐商品化，但底层数据与存储基础设施仍是长期扩展的决定性基石。 - [22:48](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1368s) 与其盲目堆砌数百个分散 Agent，不如聚焦少数核心超级 Agent，并设立统领红线的‘共享伦理 Agent’及专职负责审计与权力制衡的治理 Agent。 - [26:06](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1566s) 智能体的终极定位是人机协同（Human-in-the-Loop），其使命不是直接取代人类，而是剥除易错与重复劳动，赋能一线员工并推动全公司的组织级转型。 - [28:14](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1694s) 企业降本增效优先从削减离岸处理和外部生成等第三方外包合同切入，通过 AI 智能体直接补齐内部员工技能盲区，把节省资金回流产品创新。 - [31:49](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1909s) 端侧小模型（SLM）与开放互操作标准正加速普及；混合部署虽兼顾隐私与成本，但高度依赖本地软件管理运维栈，同时需要全面复兴对海量冷数据的治理能力。 - [35:27](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=2127s) 穿透行业营销迷雾需要结合一线客户沟通、开源社区交流、AI 竞品检索工具以及高频生态产品 Demo，并通过紧密的同行专家网络进行交叉验证去噪。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=4s) 嘉宾介绍与破冰趣事 圆桌讨论在轻松幽默的氛围中拉开序幕。主持人兼 TIAA（美国教师退休基金会）高级常务董事兼首席架构师 Geeta Pyne 首先作自我介绍。她指出 TIAA 是一家拥有百年历史、致力于保障美国人养老与财富管理的机构。Geeta 调侃称自己虽然头衔是首席架构师，但实际上更像是一个“首席洗瓶工（Chief Bottle Washer）”或“保洁员”，职责就是在架构、工程与数据层面收拾烂摊子、理清头绪。在邀请各位嘉宾介绍自己的同时，她提出了一个破冰要求：每人必须分享一件关于自己的趣事。 Cox Automotive 人工智能与机器学习副总裁 Jodi Blomberg 率先接话。她介绍 Cox Automotive 涵盖软件业务以及汽车拍卖、技师维修等技术驱动型服务；自己此前曾就职于 Waste Management 和 Charles Schwab，单纯出于对各类行业 AI 应用的热爱。谈及趣事时，Jodi 幽默地陷入自我纠结，表示自己本来觉得自己挺搞笑的，但临场却想不出段子，被 Geeta 调侃这就像“棕色的狗就是一只棕色的狗”一样无厘头。随后，Cox Communications 负责 AI 基础设施与扩展、IT 运营及公私有云的副总裁助理 Matthew Shorts 作了介绍。他将趣事转化为冷知识，透露自己曾经尝试过“徒手抓鲶鱼（Noodling）”，引得现场观众纷纷准备掏出手机搜索。 紧接着，Hyland Software 首席技术官 Tim McIntire 进行了自我介绍。他拥有在 Teradata、斯克里普斯海洋研究所（Scripps Oceanography）、高性能计算及神经网络领域深耕数十年的资深数据背景，并提到拥有 30 年企业内容管理历史的 Hyland 正迎来巨大机遇；他分享的趣事则是自己今天特意穿了一双亮粉色鞋底的鞋子。最后，来自 Hammerspace 的 AI 解决方案高级总监 Samuel Newnam 介绍了自己 25 年来在存储、数据中心虚拟化及网络领域的背景，强调自己专注于将产品架构与 AI Agent、MCP（Model Context Protocol）服务紧密结合。他的趣事则是一个典型的“极客冷笑话”——他把所有老爹笑话（Dad jokes）都存在数据库里，以便需要时能以毫秒级速度检索调用。 - **Geeta Pyne**：TIAA 高级常务董事兼首席架构师，拥有 30 年跨行业架构与数据经验。 - **Jodi Blomberg**：Cox Automotive AI/ML 副总裁，专注汽车拍卖与维修等软件及技术赋能场景。 - **Matthew Shorts**：Cox Communications AI 基础设施、IT 运营与混合云副总裁助理。 - **Tim McIntire**：Hyland Software 首席技术官，深耕企业级非结构化数据与 HPC。 - **Samuel Newnam**：Hammerspace AI 解决方案高级总监，专注 AI 数据架构、Agent 与 MCP 服务集成。 --- ### [03:55](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=235s) Agent 规模化落地的现实：经济账与框架幻象 谈话切入正题后，Geeta 引述了当天上午主题演讲的内容，向嘉宾抛出核心问题：当前业界究竟是在构建什么形态的 Agent？从早期的概念验证（PoC）走向企业级规模化部署的过程中，基础设施和成本面临怎样的真实挑战？ Jodi Blomberg 率先表达了审慎且务实的观点。她更倾向于使用“智能体系统（Agentic Systems）”而非孤立的单个“Agent”来描述当前的架构演进。Jodi 坦言，她对目前许多大型企业是否真正实现了超大规模的“智能体网络（Agentic Mesh）”持怀疑态度，因为企业背后的经济模型尚未跑通。构建一个智能体系统非常容易，但系统运行所消耗的 Token 成本极易超过其创造的商业价值。她直言：“如果你声称自己已经完全掌握了如何从成本和基础设施两个维度来规模化 Agent，那你可能多少有点‘自欺欺人（a little bit delusional）’。” 在 Jodi 看来，当前企业在智能体规模化层面的真实策略是“战略下注”。在探索正确答案的阶段，团队愿意多消耗一些 Token 来验证系统闭环；而一旦探索出可行路径，就会回归传统机器学习的打法——提炼出可复现的最佳模式，并将其固化与持续推广。 --- ### [06:44](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=404s) 重新定义 Agent 与 Hyland 的联邦数据基座 针对业内频繁将传统自动化脚本包装为 Agent 的现象，Geeta 提出应厘清“自动化”与“真正的 Agent/Agentic”之间的本质区别，并邀请 Tim McIntire 给出定义。 Tim 分享道，在近期参加的 AI 开发者大会上，4 小时的主题演讲光是定义 Agent 就耗费了整整 2 个小时。在他看来，真正意义上的 Agentic 工作流必须具备三大核心特征：主要运行方式是**自主的（Autonomous）**、执行过程是**持续或近乎持续的（Continuous / Near-continuously）**，并且能够在无需人类实时介入的情况下提供**计算智能（Compute Intelligence）**。 谈及 Hyland 在该领域的实践，Tim 指出，企业中 80% 的数据都是非结构化数据，而这正是 Hyland 深耕 30 年的内容管理核心领域。Hyland 近期推出了 Agent 构建框架（Agent Builder Framework）的测试版。他强调，构建企业级智能体平台最关键的壁垒并非模型本身，而在于如何以受控、可理解且合规的方式，跨多个记录系统（Systems of Record）实现对海量联邦数据（Federated Data）的安全访问。 ```mermaid flowchart TD accTitle: Enterprise Agentic System and Federated Data Architecture subgraph S1["底层数据层 (Enterprise Data Layer)"] D1["非结构化内容 (文档/元数据)"] D2["多个核心记录系统 (Systems of Record)"] D3["多源异构存储 (跨多云/本地)"] end subgraph S2["数据治理与联邦抽象 (Hyland Platform)"] F1["联邦数据访问与安全管控"] F2["统一上下文与元数据索引"] end subgraph S3["智能体工作流 (Agentic Framework)"] A1["Agent Builder Framework"] A2["自主与持续执行引擎"] A3["免人工介入的智能决策"] end D1 --> F1 D2 --> F1 D3 --> F1 F1 --> F2 F2 --> A1 A1 --> A2 A2 --> A3 ``` --- ### [08:28](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=508s) 数据混乱治理、全局命名空间与基础设施级 Agent 面对 Tim 提到的数据难题，Geeta 将话筒交给专注于数据架构的 Hammerspace 专家 Samuel Newnam，探讨底层存储与架构如何应对 Agent 带来的变革。 Sam 揭示了企业数据现状中普遍存在的“混沌状态”：数以 PB 计的数据散落在各个孤岛，拥有各自独立的认证与登录体系。以往，企业面临的挑战是一个 26 岁的数据科学家进门并索要系统全部权限的“上帝模式（God access）”，至少这背后还是一个具体的人，有可追踪的面孔和身份；但当演进至 Agent 时代，系统交互变成了“计算机与计算机直接对话”，权限审计、监控追踪以及全链路可观测性（Observability）变得空前复杂。 为了化解这一难题，Hammerspace 倡导通过构建“全局命名空间（Global Namespace）”来简化、自动化并加速数据流动。随着物理 AI（Physical AI）与多模态智能体的发展，计算不再局限于单一数据中心，而是受制于“GPU 引力”与“电力引力（Power Gravity）”。数据必须突破协议、供应商与地理位置的限制，为分布在各处的 Agent 提供统一、安全的访问端点。同时，Sam 透露 Hammerspace 也在践行“喝自己的香槟（Drinking our own champagne）”——率先将自主 Agent 部署到自身基础设施内部，用于实时监控、分析并优化智能体工作流中的端到端延迟。 - **数据权限与可观测性跃迁**：从人类用户的“上帝权限”管理转向机器与机器通信（M2M）的实时监控与细粒度追踪。 - **全局命名空间（Global Namespace）**：跨越不同地理区域、云厂商和底层存储协议，将分散的 PB 级数据整合为统一逻辑访问点。 - **算力引力迁移**：从传统的单体存储架构转向跟随 GPU 引力与电力引力分布的分布式数据基础设施。 - **基础设施内生智能**：优先在基础设施层引入 Agent，主动诊断与优化 Agentic 复杂链路中的网络及存储延迟。 ### [10:04](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=604s) 联邦数据平台的延迟监控与规模化现状的剖析 在讨论智能体工作流时，可观测性与底层数据平台的协同至关重要。企业需要通过 Zipkin 等链路追踪工具到各类元数据统计系统（MD stats），全面洞察智能体工作流中的每一个环节。这不仅能有效加速检索增强生成（RAG）和智能体任务的流转，更能精准定位系统延迟发生的具体节点。例如，当数据被重定位到不同区域、或是从边缘端迁移到云端时，系统能评估这对性能的潜在帮助以及随之而来的成本指标变化。通过这种联邦化的多系统互通方式，企业致力于构建一个真正支撑多样化负载的底层数据平台。 针对智能体规模化（Scaling Agents）的真实进展，嘉宾们直言不讳地指出了当前行业中的普遍误解。所谓的大规模落地往往被夸大，很多宣称“部署了 100 个智能体”的企业，拆解到底层其实只是十来个真正的核心智能体，下方挂载了大量细碎的基础工具代理（utility agents）。真正的破局点在于建立清晰的本体（ontology），将细粒度、可复用的架构组件组合成具备层级结构的集群能力，形成可组合的架构（composable architecture）。当前展厅中充斥着各类智能体构建平台，竞争异常碎片化，因此企业更倾向于依靠“编排智能体”（orchestrator agent，而非所谓“母体智能体”）跨领域调度协同，通过在全公司范围内解锁微型用例来摸索标准化路径，而非盲目进行大规模押注。 ```mermaid flowchart TD accTitle: 智能体编排与可组合架构 accDescription: 展示编排智能体如何调度底层细粒度工具智能体并基于联邦数据平台运行 subgraph Orchestration["智能体协同与控制层"] OA["编排智能体 (Orchestrator Agent)"] Ontology["智能体本体与层级结构 (Ontology)"] OA --- Ontology end subgraph UtilityLayer["细粒度执行单元"] UA1["专用智能体 A"] UA2["专用智能体 B"] UA3["工具型代理 (Utility Agents)"] end subgraph DataInfra["联邦数据与可观测性平台"] Obs["Zipkin / 性能监控与延迟追踪"] FedData["跨多云/边缘的联邦数据访问"] end OA --> UA1 OA --> UA2 OA --> UA3 UA1 & UA2 & UA3 --> DataInfra ``` - **链路延迟定位**：通过跨端追踪分析数据重定位对响应时间和云/边缘成本的影响。 - **智能体架构解构**：避免将简单的辅助型小工具夸大为独立的大型智能体。 - **渐进式试点策略**：不盲目推进大额押注，而是利用编排器（Orchestrator）在各部门验证小规模可复用能力。 --- ### [12:02](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=722s) 企业 AI 素养普及与自由创新和标准化的博弈 推广智能体不能脱离底层基本功的建设，而推动管理层与业务部门理解这一逻辑的前提是全员 AI 素养的提升。有企业在去年年底意识到自身在基础能力上的短板，果断对组织架构进行了重组与重置，以匹配行业技术的演进节奏。为了让认知渗透到业务各层级，企业在不同的技术栈与业务部门中设立了“布道者”（champions），由他们结合专业业务场景推动认知升级。实际推行中，基础 AI 素养培训的紧迫性远超预期，甚至需要从最基础的认知纠偏做起——向同事解释清楚“ChatGPT 并不是一个美化版的 Google 搜索引擎”。所幸自上而下的推动获得了高管层的全力支持，管理层明确要求加速行动，为基层探索提供了空间。 在鼓励提速的过程中，企业面临着速度、实验与标准化之间的艰难平衡。当团队全速向前奔跑时，管理层必须决定给予各组多少自由试错的“游乐场”（playgrounds），以及何时收拢建立统一平台。过度发散会导致“同一件事用二十种不同方式去做”，而过早限制又会压制创新。尤其是在涉及数据副本管理（copy data management）、安全审计、红蓝军对抗（red teams and blue teams）等合规要求时，必须在受控沙盒中推进，在智能体走向规模化之前先把运营、治理与运行机制彻底厘清。 --- ### [15:08](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=908s) 试错实践：“伤疤”作为内部能力与企业三大落地场景 谈及在试错过程中留下的“伤疤”（scar tissues），企业将其视为转化为组织核心肌肉记忆的必经之路。以 Cox Automotive 为例，其智能体落地实践清晰地划分为三大类场景： 1. **赋能自身研发**：重塑软件开发本身的流程与工程效率。 2. **赋能内部业务与现场人员**：服务线下拍卖场的工作人员或现场技术员，帮助他们处理日常文书工作、优化繁琐的物理操作工作流。 3. **面向客户的外部商业产品**：打造基于智能体的商业化解决方案推向市场。 企业目前的核心打法是将前两个内部场景中踩坑积累的经验，反哺并简化第三类面向外部客户的产品交付。在多云及混合云（multicloud hybrid cloud）环境中，跨生态调度数据常常会踩坑——例如曾因配置不当在昂贵的数据管道上输送了过量数据而产生巨额账单。但正因为这些错误是在内部流程中发生的，风险被严格隔离在企业内部，未外溢给客户。在知识管理（knowledge management）等直面客户的应用上，即使初次架构设计出现偏差，团队也能迅速重构，把内部沉淀的“伤疤”转化为自信的架构认知，更安全地为最终用户释放价值。 | 智能体应用维度 | 目标受众与典型任务 | 关键策略与价值闭环 | |:--- |:--- |:--- | | **自身研发流程** | 内部开发与工程团队，重构代码构建与系统交付流程 | 率先试错，摸清多云网络与计算开销边界 | | **内部现场业务** | 拍卖场工作人员、现场技术员，辅助文档归档与日常工作流优化 | 降低重复性文书负担，验证沙盒安全与数据管控 | | **外部商业产品** | 企业外部最终客户，交付融合企业上下文的智能体服务 | 继承内部沉淀的最佳实践，避免客户面临架构与成本风险 | ### [19:05](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1145s) 上下文工程与 Agent 时代的安全与数据暴露隐患 Tim McIntire 探讨了“企业上下文引擎”（Enterprise Context Engine）的构想——它作为一个运行在平台底层的 Agent 工作流，不断将企业私有数据转化为上下文，无缝供给由外部 LLM 驱动的上层应用。这种模式超越了早期的提示词工程（Prompt Engineering）、传统 RAG 或手工上下文工程，让 Agent 自动化管理数据注入过程。 主持人 Geeta Pyne 迅速提议稍后对“上下文工程”进行深入展开。随后，Samuel Newnam 谈及企业在落地 Agent 时面临的现实创伤（Scars）。他指出，“数据”与“数据泄露”往往形影不离。许多企业客户最深层的恐惧在于：一旦启用自主 Agent，究竟应该赋予它们多大的权限？Agent 本身构成了全新的攻击向量。尤其是像模型上下文协议（MCP）这样新兴的协议与服务平台，虽然赋予了模型强大的外部工具调用能力，但整个行业在安全标准制定上仍处于起步阶段。在 GPU 高速吞吐通信的同时，企业如何确保两个互联系统不会凭空创建敏感数据，或是越权访问原本受限的企业资产？这使得 Agent 的权限边界设定、自主自由度控制以及可观测性（Observability）成为了当下最棘手的核心议题。 ```mermaid flowchart TD accTitle: Enterprise Agent Security and Context Engine Architecture subgraph EnterpriseContext["企业上下文引擎 (Enterprise Context Engine)"] Data["企业私有数据 (Enterprise Data)"] --> AgenticContext["Agent 上下文工作流"] end subgraph SecurityGate["安全与治理边界 (Security & Observability)"] AgenticContext --> AuthCheck{"权限与合规检查\n(MCP / 访问控制)"} AuthCheck -- 授权通过 --> SafeContext["结构化安全上下文"] AuthCheck -- 违规阻断 --> Log["安全日志与可观测审计"] end SafeContext --> LLMApp["外部 LLM 驱动的应用"] ``` --- ### [21:05](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1265s) 规模化（Scaling）的真正含义：基础设施选型、成本与 ROI 面对“如何在企业中实现 Agent 规模化扩展”的问题，嘉宾们从基础设施与经济模型两个维度展开了解剖： * **底层基础设施的架构抉择**：Samuel Newnam 认为企业当前正处于关键十字路口。过去存放在传统 NAS 存储上的非结构化数据在性能上已无法满足 AI 需求。企业高管在技术选型上面临“规模化瘫痪”（Scale Paralyzed）：是选择开箱即用的闭源一体化系统，还是自行拼装开源组件？算力究竟部署在边缘端、公有云，还是去各种 Neo-Cloud 争夺 GPU 算力时长？他强调，尽管算力正在逐渐商品化（Commodity），但底层数据与存储平台依然是决定 Agent 能否在 2026 年、2027 年乃至 2028 年持续扩展的真正基石。 * **性价比与 ROI 的核心约束**：Tim McIntire 补充道，尽管业界对大规模数据管理已相对熟练，但当前 AI 与 Agent 规模化扩展的核心难点始终围绕着“性价比”（Price Performance）与“投资回报率”（ROI）。企业必须为具体场景匹配最合适、最具成本效益的模型，否则 Agent 工作流消耗的推理成本将迅速吞噬业务产出的价值。 --- ### [22:48](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1368s) 横向扩展 vs 纵向渗透：Agent 架构的拓扑与伦理中枢 Matthew Shorts 提出了思考 Agent 规模化的两个核心维度：横向扩展（引入更多不同功能的 Agent）与纵向深化（将单一通用 Agent/母 Agent 深度渗透至整个组织的各个业务环节）。如果一个具备编排能力的“母 Agent”（Mother Agent）能够跨部门创造稳定价值，企业应当集中资源最大化压榨该 Agent 的 ROI，而不是盲目在系统里堆积数百个各自为政的细分 Agent。 Jodi Blomberg 赞同这一观点，并引用了零售巨头沃尔玛（Walmart）的案例：沃尔玛在尝试了大量分散 Agent 后，选择收缩并聚焦于少数核心超级 Agent（如客服 Agent、销售 Agent）。Jodi 提出，与其争论是搞“母 Agent”还是细分 Agent，企业更需要一个“共享伦理/价值观 Agent”（Shared Ethics Agent）。无论底层的 Agent 工作流如何编排，都必须统一运行在企业既定的伦理红线与核心价值观约束之下。 Geeta Pyne 则从软件工程视角总结了 Agent 的分类体系： 1. **业务领域 Agent**：按照企业的业务域或功能分类法（Capability Taxonomy）划分，天然继承对应领域的数据权限与访问控制。 2. **技术与治理 Agent**：常驻运行的底层治理节点，专门负责可观测性（Observability）、审计、伦理校验和权力制衡，确保系统“不仅具备做某件事的技术能力，更符合做这件事的合规标准”。 --- ### [26:06](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1566s) 从系统到组织：人机协同与以人为本的组织转型 当讨论转向“企业如何在 Agent 落地过程中赋能员工”时，专家们一致驳斥了“完全取代人类”的激进论调： * **人机协同（Human-in-the-Loop）**：Tim McIntire 指出，在超过 90% 的企业业务流中，Agent 的目标绝不是踢开人类，而是通过提供更全面、精准的上下文信息，让员工更高效地做出决策。 * **剔除易错与重复劳动**：Matthew Shorts 进一步校准了这一逻辑：Agent 的真正使命是剥除人类在流程中引入的失误和错误。将员工从枯燥、重复且容易出错的任务中彻底解放出来，使他们能够将精力投入到更具创新性、更高价值的业务创造中，这也能极大提升员工对工作的热情与成就感。 * **全员转型与现场痛点自动化**：Jodi Blomberg 分享了 Cox Automotive 的现场调研经历——团队直接跟车陪同外勤一线人员，当面询问“你日常工作中最讨厌哪一部分？”，而那一部分往往就是技术团队最优先通过 Agent 实现自动化的痛点。她援引公司首席产品官（CPO）的话指出，Agent 落地绝非单纯的技术上线，而是一场“全公司的组织级转型”（Whole-Company Transformation），无法仅靠几场简单的内部午餐分享会（Brown Bag Session）轻松过关，必须从高层战略到底层执行双向推进。 ### [28:14](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1694s) 成本压力与劳动力替代：AI 是“助手”而非“替代者” 主持人抛出了一个尖锐的话题，直指来自董事会或 CEO 的实际压力：企业在部署智能体（Agent）时，究竟追求多大幅度的成本节约？这是否意味着直接缩减员工编制（FTE）？针对这一“触碰烫手山芋”的问题，Jodi Blomberg 率先作答。她透露，其所在团队设立的 AI 加速器在今年前六个月内就构建了 27 个概念验证（PoC），并挖掘出了约 2000 万美元的效率提升空间。但她明确指出，这些降本增效目前主要来自于削减外部供应商开支，而非解雇内部员工——例如取消外包内容生成的合同，或是停用负责手动读取 PDF 的离岸外包公司。先从这些低摩擦、低压力的外部环节切入，不仅能直接增厚企业利润，还能保护内部员工的积极性，为团队去开发真正能商业化收费的创新产品争取空间。 随后其他嘉宾补充指出，AI 智能体为企业带来了一种全新的“专业能力引入”模式。过去企业若缺乏某种高度垂直的专业技能，通常只能通过高薪雇佣全职员工或采购昂贵的第三方专业服务来弥补；而现在，企业可以通过智能体直接赋能现有员工，为其配备此前不具备的技能，从而在不增加编制、不裁撤团队的前提下实现业务能力的跃升。这种模式直接冲击的是第三方外包合同，实现了“把节省下来的每一分钱重新投回产品创新”的良性循环。 - **AI 加速器成效**：半年内构建 27 个 PoC，识别出约 2000 万美元的效率优化空间。 - **降本首选目标**：优先替代离岸 PDF 处理和外部内容生成等第三方外包合同。 - **组织能力重塑**：通过 AI 智能体直接补充内部员工技能盲区，无需扩充编制即可引入垂直专业能力。 --- ### [30:27](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1827s) 研发预算、快速试错与人机协同的心理机制 Samuel Newnam 探讨了企业在引入智能体工作流时的预算架构与心理预期。他强调，管理层必须意识到 AI 领域的研发（R&D）依然遵循严谨的科学探索规律，不能因为一切都在云端和算法模型中运行，就忽视了试错所需的时间与资金成本。企业一方面需要设立专属的探索预算，另一方面也必须建立严格的投资回报率（ROI）评估机制，确保团队能够“快速试错”（Fail Fast），而不是无休止地进行无效实验。 在人机交互层面，Samuel Newnam 幽默地调侃道：“下次跳槽谈判时，我要求的福利可能不是配一个行政助理，而是配一个专属 AI 智能体。”他指出，从零售、娱乐到企业日常运营，数字化身（Digital Human）与数字伴侣正在重构人机交互形态。这种变革带给员工的不仅是效率数字，更是心理上的“安全感”与“确定性”——通过 AI 的实时校验与辅助，员工可以消除工作中犯错的恐惧，提升交付质量与职业成就感，这种心态维度的价值往往是传统财务报表无法直接量化的。 ```mermaid flowchart TD accTitle: 智能体企业部署与价值闭环 subgraph Budget["研发与试错机制"] A["设立专属探索预算"] --> B["快速试错 (Fail Fast) 机制"] B --> C["ROI 与效率评估"] end subgraph Execution["智能体落地场景"] C --> D["削减第三方外包服务"] C --> E["赋能内部员工专业技能"] end subgraph Impact["双重业务价值"] D --> F["直接节约资金回流创新"] E --> G["提升员工交付质量与心理安全感"] end ``` --- ### [31:49](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1909s) 智能体规模化面临的架构趋势：端侧小模型、互操作标准与混合部署 当谈到支撑智能体大规模落地必须发生的技术创新时，嘉宾们提到了几个核心技术风向。首先是部署在边缘端的**小型语言模型（SLM）**，因其轻量、低延迟和响应迅速而备受关注；其次是**互操作性标准**的普及，业内正在摆脱过去打补丁式的割裂连接方式，积极拥抱如模型上下文协议（Model Context Protocol, MCP）以及 Google 相关的互联规范，以构建开放协同的智能体生态。 在算力与基础设施层面，**混合部署（Hybrid Deployment）**由于兼顾了极致的性价比与数据隐私要求，正从曾经复杂的“科研玩具”演进为可落地的企业级架构。企业既可以选择在本地数据中心构建全栈架构（例如依托 NVIDIA Dell AI 工厂及 NIM 微服务栈紧邻企业数据仓库运行），也可以采用“本地数据 + 云端向量化/算力”的分离架构。然而，混合部署的最大瓶颈不仅在于获取充足的 GPU 硬件配额，更在于**本地软件管理栈（Management Stack）**的运维复杂度——如果无法高效管理本地部署的软件堆栈，其性价比优势很快会被高昂的运维成本吞噬。 与算力并行的是**数据管理（Data Management）**的全面复兴。随着智能体以远超人类打字数百万倍的算力吞吐数据，过去被长期堆在 SharePoint 角落、版本混乱的“冷数据”必须被重新清洗与治理。预计到 2027 年全球将产生高达 300 ZB 的海量数据，如何在如此庞大的数据洪流中厘清数据资产价值并建立严格的访问与调度机制，是实现智能体规模化扩展的决定性前提。 - **端侧 SLM 与开放互联**：边缘小模型配合 MCP 等互操作标准，打破应用间的数据孤岛。 - **混合部署范式**： - *全本地部署*：本地数据仓库配合 NVIDIA NIM 软件栈与 AI 工厂硬件。 - *拆分架构*：核心敏感数据留存本地，向量化与模型推理算力上云。 - **数据治理新挑战**：从传统的“数据堆砌”转向智能体级的高频检索治理，应对 2027 年 300 Zettabytes 级别的全球数据规模。 --- ### [35:27](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=2127s) 现场问答：穿透行业炒作与获取一手真实信息 在现场提问环节，一位自称“极度怀疑论者”的观众提问：在充满营销夸大与夸张宣传的 AI 行业中，专家们究竟通过哪些渠道获取客观真实的信息，如何保持对技术落地可行性、成本与能耗的独立思考？ Samuel Newnam 分享道，获取第一手真相的最有效途径其实非常朴素——就是与一线客户保持高频的直接沟通。他在每天密集的 Zoom 会议中倾听各类企业对智能体的不同定义、实际遇到的数据脏乱差问题与落地挫折；此外，开源社区交流、行业展台面对面讨论以及同行间的非公开交流，都远比那些由赞助商包装的播客或商业广告更加真实客观。当主持人打趣追问“你们现在还给 Gartner 付费买咨询报告吗”时，嘉宾笑着坦承行业内依然存在这种“付费参与”的商业机制，但真正有价值的技术认知早已来自于开源生态、一线实践以及智能体辅助的高效信息挖掘之中。 ### [37:04](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=2224s) 穿透营销迷雾：通过 AI 搜索与密集 Demo 保持技术敏锐度 在探讨如何紧跟日新月异的 AI 趋势时，嘉宾们分享了各自获取前沿信息与进行竞品调研的实用方法。利用像 Perplexity 这样的 AI 检索工具，可以直接对各类产品进行深入的竞品分析（Competitive Analysis），从而看清底层技术与真实能力的本质。这种方式能让技术团队快速看透厂商包装出来的营销话术，暴露出产品真正具备的“真实性”（Authenticity），让浮夸的宣传无所遁形。除了依赖 AI 搜索工具，多与行业内的人交流也是不可替代的信息源，例如收听《AI Daily》等行业知名播客以获取灵感。 与生态伙伴的实际交流和产品演示（Demo）同样是把握技术脉搏的关键。有嘉宾坦言自己的团队其实“苦不堪言”，因为他非常热衷于引入潜在合作伙伴安排产品 Demo。团队成员常常抱怨并翻白眼说：“我们已经有这个功能了，或者我们已经做了类似的东西。”但他总是坚持认为，外部伙伴可能在某些独特的维度上提出了团队从未考虑过的解法。除了积极约见合作伙伴看 Demo 之外，订阅《The Rundown》等行业快讯也是极好的习惯，它每天提供精炼的信息切片，让技术管理者在早上 9 点就能迅速补齐前一天傍晚到夜间数小时内 AI 领域发生的最新突发动态。 - **调研与检索工具**：借助 Perplexity 等新型 AI 检索产品直接透视竞品实力，剥离表层营销宣传。 - **Demo 驱动洞察**：顶住内部团队“已有类似方案”的吐槽，通过高频观看合作伙伴的产品演示挖掘盲区创新。 - **资讯追踪渠道**：日常收听《AI Daily》播客，并利用《The Rundown》晨间简报快速掌握过去数小时的行业突变。 --- ### [38:14](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=2294s) 过滤行业噪音：硬核白皮书、精选 Newsletter 与“场外求助”同行验证 针对更深层次的技术吸收，嘉宾们展现了不同的学习路径。有嘉宾给自己设定了“每周硬核阅读一篇技术白皮书（White Paper）”的挑战，并幽默地表示自己就是纯粹“喜欢里面的数学推导部分”。这一硬核习惯立即引发了现场其他嘉宾的打趣与调侃：“居然真有人读那些白皮书啊？”引来会心一笑的回应：“对，我们当中的一部分人确实会读，不过在一堆论文里挑选具体读哪一篇其实相当困难。” ```mermaid flowchart TD accTitle: 跟踪 AI 前沿与去伪存真的信息流架构 subgraph S1["信息输入源 (Information Sources)"] A["学术白皮书与数学论文<br/>(Weekly White Papers)"] B["每日行业快讯<br/>(The Rundown / The Turing Post)"] C["生态 Demo 与 AI 检索<br/>(Perplexity / Partner Demos)"] end subgraph S2["辨别与过滤 (Filter & Triage)"] D["梳理事实脉络<br/>(Piece together what is true)"] end subgraph S3["同行交叉验证 (Peer Validation)"] E["'求助场外好友' 机制<br/>(Phone a Friend)"] F["同行技术质询与去噪<br/>('Hey, that doesn't seem true')"] end A --> D B --> D C --> D D --> E E --> F ``` 除了深啃白皮书，日常收件箱里订阅的高质量通讯（如《The Turing Post》等 AI Newsletter）也是构建全局认知的重要拼图。面对铺天盖地真假难辨的技术炒作与突破新闻，嘉宾总结出了一套“拼图辨真伪”的务实策略：先通过每日快讯碎片拼凑出整体脉络，遇到存疑或看似不合理的技术宣称时，直接启动“场外求助热线（Phone a friend）”，向行业同行与专家好友发消息求证：“嘿，这说法看起来不太靠谱，你觉得呢？”通过紧密的同行网络完成最后一步的去伪存真。 ## Notable quotes > [03:55](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=235s) 如果你声称自己已经完全掌握了如何从成本和基础设施两个维度来规模化 Agent，那你可能多少有点‘自欺欺人’。 — Jodi Blomberg > [06:44](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=404s) 真正意义上的 Agentic 工作流必须具备三大核心特征：主要运行方式是自主的、执行过程是持续或近乎持续的，并且能够在无需人类实时介入的情况下提供计算智能。 — Tim McIntire > [12:02](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=722s) ChatGPT 并不是一个美化版的 Google 搜索引擎。 > [19:05](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1145s) 数据与数据泄露往往形影不离。 — Samuel Newnam > [22:48](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1368s) 与其争论是搞‘母 Agent’还是细分 Agent，企业更需要一个‘共享伦理/价值观 Agent’。无论底层的 Agent 工作流如何编排，都必须统一运行在企业既定的伦理红线与核心价值观约束之下。 — Jodi Blomberg > [26:06](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1566s) 在超过 90% 的企业业务流中，Agent 的目标绝不是踢开人类，而是通过提供更全面、精准的上下文信息，让员工更高效地做出决策。 — Tim McIntire > [26:06](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1566s) Agent 的真正使命是剥除人类在流程中引入的失误和错误。将员工从枯燥、重复且容易出错的任务中彻底解放出来。 — Matthew Shorts > [26:06](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1566s) Agent 落地绝非单纯的技术上线，而是一场‘全公司的组织级转型’，无法仅靠几场简单的内部午餐分享会轻松过关。 — Jodi Blomberg > [30:27](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=1827s) 下次跳槽谈判时，我要求的福利可能不是配一个行政助理，而是配一个专属 AI 智能体。 — Samuel Newnam > [38:14](https://www.youtube.com/watch?v=ORk6e_H9FWo&t=2294s) 嘿，这说法看起来不太靠谱，你觉得呢？ — Samuel Newnam ## Mentioned Geeta Pyne, Jodi Blomberg, Matthew Shorts, Tim McIntire, Samuel Newnam, TIAA, Cox Automotive, Cox Communications, Hyland Software, Hammerspace, Waste Management, Charles Schwab, Teradata, Scripps Oceanography, Walmart, NVIDIA, Dell, Gartner, Google, Perplexity, ChatGPT, Zipkin, Model Context Protocol (MCP), NVIDIA NIM, SharePoint, AI Daily, The Rundown, The Turing Post ## Video description Samuel Newnam, Sr. Director of AI Solutions, Hammerspace Jodi Blomberg, VP, AI and Machine Learning, Cox Automotive Tim McIntire, EVP, Chief Technology Officer, Hyland Matthew Shorts, AVP, AI, IT Operations, and Public and Private Cloud, Cox Communications Geeta Pyne, Senior Managing Director and Chief Architect, TIAA Abstracts AI agents have rapidly evolved from experimental tools to transformative forces capable of delivering significant business value across enterprise functions. This panel will explore the critical infrastructure components—hardware, software, and human resources—that companies must deploy to effectively scale AI agents throughout their operations. Panelists will discuss essential digital tools and platforms that enable seamless agent integration, management, and optimization within complex enterprise ecosystems. Topics will include data requirements, compute resource allocation, and strategies for maintaining robust performance as agent complexity increases. Attendees will gain insights into best practices for building scalable, efficient, and adaptable AI infrastructures designed to unlock the full potential of enterprise-wide agent deployments. [Source](https://www.youtube.com/watch?v=ORk6e_H9FWo)

[3:55] Jodi Blomberg: 如果你声称自己已经完全掌握了如何从成本和基础设施两个维度来规模化 Agent，那你可能多少有点‘自欺欺人’。 [6:44] Tim McIntire: 真正意义上的 Agentic 工作流必须具备三大核心特征：主要运行方式是自主的、执行过程是持续或近乎持续的，并且能够在无需人类实时介入的情况下提供计算智能。 [12:02] ChatGPT 并不是一个美化版的 Google 搜索引擎。 [19:05] Samuel Newnam: 数据与数据泄露往往形影不离。 [22:48] Jodi Blomberg: 与其争论是搞‘母 Agent’还是细分 Agent，企业更需要一个‘共享伦理/价值观 Agent’。无论底层的 Agent 工作流如何编排，都必须统一运行在企业既定的伦理红线与核心价值观约束之下。 [26:06] Tim McIntire: 在超过 90% 的企业业务流中，Agent 的目标绝不是踢开人类，而是通过提供更全面、精准的上下文信息，让员工更高效地做出决策。 [26:06] Matthew Shorts: Agent 的真正使命是剥除人类在流程中引入的失误和错误。将员工从枯燥、重复且容易出错的任务中彻底解放出来。 [26:06] Jodi Blomberg: Agent 落地绝非单纯的技术上线，而是一场‘全公司的组织级转型’，无法仅靠几场简单的内部午餐分享会轻松过关。 [30:27] Samuel Newnam: 下次跳槽谈判时，我要求的福利可能不是配一个行政助理，而是配一个专属 AI 智能体。 [38:14] Samuel Newnam: 嘿，这说法看起来不太靠谱，你觉得呢？

[0:04] Jodi Blomberg: apparently I just really like AI, but I like all the industries. [0:04] Sam Newnam: the real key to all the dad jokes are to keep them in a database so you can get to them real easy. [4:00] Jodi Blomberg: none of us want to define an agent [4:00] Jodi Blomberg: it's deceptively easy to build an agentic system that costs more to run than the money you're making off of it. [4:00] Jodi Blomberg: probably a little bit delusional [10:41] Matthew Shorts: every aisle has like 20 agent builder platforms [12:02] Matthew Shorts: do more of this and you need to do it faster [15:52] Jodi Blomberg: You should wear it proudly because you learned something, you didn't do it twice. [23:43] Jodi Blomberg: just because you can, doesn't mean you should. [28:14] touch the hot stove. [34:16] clean your room [37:04] We already have that or we already have something like that [38:14] People read those?

[0:04] Good afternoon everyone. How is everyone [0:07] doing? [0:09] >> Yes. Cool. So my name is Gita Pine. I am [0:12] the senior managing director and chief [0:14] architect for a company called TIAA [0:16] which is teachers insurance and annuity [0:19] association. We are a 100 plus years old [0:21] company. We are in the business of [0:23] making sure Americans have enough to [0:25] retire. So retirement, wealth management [0:26] and asset management. I have been there [0:28] for about two years. I I'm I'm the chief [0:31] architect, but I say that I'm the chief [0:33] bottle washer. My job is to make sure [0:35] I'm a cleaner. So, I clean things up and [0:37] that's what I do. I have been working [0:39] for 30 years across several industry and [0:43] you know architecture, engineering, data [0:45] is my thing. So, very excited to have [0:48] this panel because extremely talented [0:51] you'll see. So, I'm going to stop [0:53] talking and pass it on to Judy to do the [0:55] introduction. [0:57] And while introding, say one thing funny [1:00] about you. [1:03] Oh, I like more warning than that. But, [1:05] um, [1:07] I'm Jody Blumbberg. I'm the VP of AI and [1:10] ML for Cox Auto. It's a collection of [1:12] software uh, businesses as well as tech- [1:16] enabled services like car auctions and [1:18] uh, technicians fixing cars. Um, and [1:21] I've been there for a couple years, not [1:23] quite. Um, and I came from Waste [1:25] Management and Charles Schwab before [1:27] that because apparently I just really [1:29] like AI, but I like all the industries. [1:32] Uh, and I do that do different things. [1:34] Um, something funny about myself. [1:41] See, I think that I'm really funny. So, [1:43] no, I can't. [1:44] >> It's just every [1:45] >> Okay, that works. [1:47] >> And that but not today, apparently. [1:49] Something funny about you. [1:51] >> Okay, [1:51] >> that was a that was your answer. I'm [1:54] funny. [1:56] >> It's like a I have worked with people [1:58] that says like a brown dog is a dog that [2:00] is brown. [2:04] >> Okay. [2:04] >> Okay. My name is Matt Schwarz. [2:08] I work for Cox Communications uh [2:10] internet service provider across kind of [2:11] the southern half of the United States. [2:13] uh my responsibility is AI um uh [2:16] infrastructure and scaling uh IT [2:18] operations and public and private cloud [2:20] for the communication side of the [2:21] business. So I am a little busy from [2:24] time to time. Something funny about me [2:26] or I would say an interesting fact. I'm [2:28] going to twist twist the question. I [2:30] have been noodling before. If anybody [2:32] knows what noodling is, about half of [2:34] you will start googling it right now. [2:38] >> Over to you. [2:38] >> Thank you. [2:40] >> Google it. Google before I answer it. [2:42] >> Google it. [2:43] >> Noodling question. Tim McIntyre. Uh, CTO [2:46] Highland Software. Um, I'm new to [2:49] Highland. Highland has been in the [2:50] enterprise content management space for [2:52] about 30 years. Uh, I've spent my career [2:54] in the enterprise data space at scale uh [2:58] at Terod Data Scripts Oceanography, high [3:00] performance computing, neural networks. [3:02] Uh, super excited to be here and super [3:04] excited to to be at Highland. And [3:06] something funny about me, I have pink [3:08] sold shoes. [3:10] Ooh, that's real funny. Thank you. [3:13] >> Nice. My uh shoes are from Amazon. Um, [3:17] but Sam Nunham, I'm the senior director [3:19] of uh AI solutions at Hammerspace. Uh, [3:22] so my job is go to market solutions, not [3:25] just product and technology, but [3:26] reference architectures, how do we [3:28] integrate with agents, MCP services, all [3:30] the fun things that uh really bring AI [3:32] alive. um been in the storage industry [3:34] for about 25 years doing everything from [3:37] data center virtualization, networking, [3:39] and uh something funny. Um I have a [3:42] couple dad jokes, but the the real key [3:44] to all the dad jokes are to keep them in [3:46] a database so so you can get to them [3:48] real easy [3:50] >> like good good you are in the right [3:52] business and in the right panel. So, so [3:55] let's talk about you know on the serious [3:58] note about AI agents and I think I would [4:00] like to start with what we heard in the [4:03] keynote this morning. So maybe each of [4:05] you as you answer the you know around [4:08] question around agents I would love to [4:10] hear are you if you're building agents [4:12] are you building mothers are you [4:13] building sisters or brothers whatever [4:15] agents you are building right so I think [4:18] I'll start with jodi you um you know we [4:22] have been seeing like we started with [4:24] pilot agents now we are seeing uh you [4:27] know lot of you know enterprise scale [4:30] deployments so you can agree or disagree [4:32] with that fact and if you do agree that [4:34] you are seeing more enterprise scale. [4:37] Talk about the demand that you know it [4:40] brings into of course with the agents on [4:42] infrastructure infra in the cost. What [4:46] is actually driving what are the kinds [4:48] of problems do you think we are really [4:50] trying to solve and give us some you [4:52] know ideas about what you are doing. [4:54] >> Well none of us want to define an agent. [4:57] No. Uh so let's start with that. Um, so [5:00] when we talk about scaling agents, I [5:02] think we talk about scaling, I'm going [5:03] to call them agentic systems rather than [5:06] agents themselves. Um, and I [5:12] somewhat doubt that many large [5:14] enterprises are scaling agents to giant [5:16] like agentic mesh proportions yet [5:18] because the economics are still not that [5:21] well understood. Um, how to use less [5:23] tokens, how to use more tokens. Um, it's [5:27] deceptively easy to build an agentic [5:28] system that costs more to run than the [5:30] money you're making off of it. Um, and [5:32] we got to figure that stuff out. Um, [5:34] that's an economics problem. I think [5:36] there's a scaffolding problem there as [5:38] well. Um, [5:41] I know I'm on a panel about scaling [5:43] agents, but I think if you were to say [5:45] that you really understood how to scale [5:47] agents both from um, cost and [5:49] infrastructure, you're probably a little [5:51] bit delusional. Like we kind of don't [5:53] know some things. Um, and so I think [5:55] when we think about scaling it is where [5:56] are we placing our bets? Um, we're [5:58] probably willing to spend a few too many [6:00] tokens to try to get to the right [6:02] answer. Um, until we get there and then [6:04] we are doing the same thing we used to [6:06] do with ML which is find a repeatable [6:09] pattern and then one that works and just [6:11] keep doing it. [6:12] >> Thank you very much. I think it you just [6:14] brought up one point around defining an [6:17] agent and you it will be great to also [6:19] like what are the properties we think [6:21] about when we talk about agent. How is [6:23] it different right from because I hear [6:26] too often just little automation that we [6:29] used to do back in the days we know all [6:31] of sometimes started to call that agent [6:34] right so it'll be good to I think define [6:36] what agentic agent means and what [6:38] agentic and when and then we'll come [6:40] back to scaling what does the different [6:42] vectors of the scaling [6:44] >> start with the definition of agent [6:47] >> whoever [6:48] >> sure so I mean there's um I was just at [6:51] the world's 's uh AI developer world's [6:55] fair. Awesome conference and I think [6:57] they spent two hours of the 4-hour [6:59] keynote defining an agent. So easy easy [7:02] topic. Um but I I believe an agent is [7:05] anything um it's a it's a system that [7:08] runs continuously uh autonomously and I [7:11] should say near continuously. So you [7:13] could have an agent that runs for a few [7:16] minutes as opposed to a few hours. But I [7:18] think true agentic workflows are [7:20] something that's primarily autonomous, [7:23] primarily continuous and provides some [7:26] amount of compute intelligence without [7:28] human intervention. [7:31] >> Awesome. So while I have you have the [7:33] mic talk about what are you doing with [7:35] agent in your Highland? [7:37] >> Yeah. So um Highland's got a really [7:40] interesting opportunity in this uh [7:42] space. So enterprise content management [7:44] is primarily unstructured data. you [7:46] know, we manage a lot of unstructured [7:48] data and process automation, metadata, [7:51] but 80% of enterprise data is is [7:53] unstructured. Um, we just released a [7:56] beta of an agent builder framework. Um, [7:59] you know, we're using the same tool [8:00] chains that I'm sure everybody up on [8:02] this uh panel is, but what we see as [8:05] really important is access to federated [8:07] data. And you know that's where Highland [8:09] really sees a place for itself in this [8:12] space is if you're going to build an [8:14] agentic platform you're going to need [8:16] access in a kind of managed [8:19] understandable way to all of your [8:21] enterprise data at scale in multiple [8:23] systems of record and that's what our [8:25] platform provides. [8:26] >> Awesome. So going to Sam regarding since [8:30] he's talking about data and you come [8:32] from a virtual database company tell us [8:34] a little bit about what Hammerspace does [8:37] and then let's get into what are you [8:39] doing with Agentic. [8:40] >> That's a great question. Um I think a [8:43] lot of people approach us because [8:44] they've got chaos right in their data [8:47] state. I mean much like you were talking [8:49] about there's pabytes of data floating [8:50] out there. It's all under different [8:52] login. Everybody's trying to do it and [8:53] you've got this 26-y old researcher [8:55] showing up wanting God access to [8:57] everything, right? And so we think about [8:59] agents at the next level, right? That [9:00] was at least a human. There was a face I [9:02] could assign to that and a permission [9:04] that person got. As we think about [9:05] agents, right? Now we've got computers [9:08] talking to computers and like where how [9:10] do you watch that? How do you monitor [9:11] that? Right? Where's the observability [9:13] in those particular things? And so our [9:15] customers are really coming to us [9:16] saying, you know, we want to simplify, [9:18] you know, automate and accelerate these [9:20] processes, right? And so the first part [9:22] of that is is let's put that together in [9:23] a global name space, right? Agents and [9:25] even as we get into physical AI, right, [9:27] are going to run everywhere. So how do [9:29] we get data everywhere, right? We used [9:30] to build these things around monolithic [9:32] storage systems and now we're dealing [9:34] with GPU gravity and really power [9:35] gravity, right? Anywhere where we can [9:36] put this stuff. So as we look at like [9:39] aentic workloads, we think how do we [9:40] feed those things with data in a single [9:42] access point, right? where there's [9:44] security, where there's understanding, [9:46] where there's, you know, data that [9:48] exceeds, you know, protocol, provider, [9:50] and geography, right? To kind of match [9:52] the needs of an organization. From an [9:54] agentic standpoint, you know, we're [9:55] starting kind of drinking our own [9:57] champagne. So, as we look at latency [9:59] within these particular workflows and [10:00] all sorts of stuff, our agents are are [10:02] looking at infrastructure first. So [10:04] we're being able to kind of understand, [10:06] hey, from Zipkin all the way into MD [10:08] stats to to anywhere within these [10:09] agentic workflows. Not only can we help, [10:12] you know, speed up rag and and agentic [10:14] workflows, but we can actually tell you [10:15] where those latencies exist. Hey, if we [10:17] relocated data to a different area, [10:19] would that potentially help this thing, [10:21] what's the cost metrics associated with [10:23] that if we move it to cloud versus some [10:25] edge solution and those type of things. [10:26] So we're really trying to deliver [10:28] metrics and kind of these, you know, [10:29] systems that have to talk to each other [10:31] in a federated way to deliver a true, [10:33] you know, data platform underneath these [10:35] workloads. [10:35] >> Awesome. Thank you so much. Matt, can [10:38] you talk about what are you doing with [10:40] scaling agents? [10:41] >> Yeah, I'm in the same boat as Jody [10:43] actually. I think there's very few [10:44] people that are truly scaling agents at [10:46] the at the rate and the pace that most [10:49] people claim. And I was in a talk [10:50] earlier, you know, there's you'll hear [10:52] someone saying, "Hey, we deployed 100 [10:54] agents." really if you decompose that [10:55] it's probably 10 really agents with like [10:58] a bunch of little utility agents [10:59] underneath it. And so from our [11:01] perspective what we're trying to do is [11:02] look at an ontology of how you would [11:05] build a um call it a cluster or [11:08] capability of agents so that you can [11:10] enable almost a composable architecture [11:13] if you will. Um so one of the challenges [11:15] is I mean especially if you walk through [11:17] the the show floor there are I think [11:20] every aisle has like 20 agent builder [11:22] platforms. Uh, and that's great. You [11:24] know, there's a there's a lot of [11:25] competition out there. It's very uh [11:27] fragmented. And so we're looking at how [11:30] we can do um very kind of granular [11:32] architecture components that are [11:34] repeatable. Um combine them into uh an [11:37] ontology that maybe even has a [11:39] hierarchy. You asked earlier about a [11:41] mother agent. Uh we hadn't used that [11:43] analogy. We think of as an orchestrator [11:45] agent. So something that can work across [11:48] other agents and bring that autonomy to [11:50] it. Um, but those are the things that [11:52] we're trying to do now. So, not making [11:53] any big bets, but unlocking a bunch of [11:56] little ones across the organization [11:57] until we start to see some [11:58] standardization in order to drive the [12:01] scaling. [12:02] >> I'll stay with you, Matt. Um, you talked [12:04] about a lot of fundamentals, right? [12:06] Whether building the ontology, being [12:09] able to like build that, you know, [12:12] understanding of your data and [12:14] orchestrator. How are you teaching this? [12:16] Like how are you educating whether it is [12:18] your board, whether it your you know [12:20] other executives that fundamentals [12:23] matter and this is what you need to have [12:26] without which agents are pretty much [12:27] useless. [12:28] >> Yeah, great question. Um I think we're a [12:31] little lucky in the sense that our [12:32] organization recognized how bad we were [12:34] at these things um late last year and we [12:37] said hey let's take a reset and a pause [12:40] and what do we need to do to restructure [12:42] our organization to match the needs of [12:44] the change of the industry and the [12:46] things that we've done. Um now the in [12:49] order to get the knowledge across the [12:51] organization we've planted champions [12:53] within the various stacks of the [12:54] organization. and those are the people [12:55] that can really push the thinking and [12:58] bring the specialization of the spaces [13:00] they work in. Um but a just a pure [13:02] literacy program across the organization [13:05] has shown to be a hu bigger need than we [13:07] actually anticipated. Um so what are how [13:10] do we say that chat GBT is not a [13:12] glorified Google which I've gotten from [13:14] peers at work. Um and so we've had to [13:16] actually reset and go back to the ground [13:19] floor of like this is what AI is and [13:22] this is what AI can do for you. But from [13:24] the top down, we've actually had some [13:26] great buyin from the executive level to [13:28] say you guys need to do more of this and [13:30] you need to do it faster. And so we've [13:31] kind of taken the reigns and run with [13:33] that as best we can. [13:35] >> Thank you. Any of you want to add to [13:36] that? [13:40] >> Such a good answer. You didn't want to [13:41] add to it. [13:42] >> Um I just say part of that is finding [13:45] the right balance too besides the kind [13:47] of uh education across the organization. [13:50] And we we're all asking uh each of our [13:53] employees, each of our engineering [13:54] groups uh to move fast, to move super [13:56] fast. And part of that is finding a [13:59] balance of when people are moving fast, [14:01] how much do we let them experiment [14:03] versus how much do we standardize, how [14:05] much do we build on a single platform. [14:07] Um, those are a lot of hard decisions [14:09] for us to make as we work with our [14:11] customers on, you know, we want to bring [14:13] this federated strategy of giving you [14:15] access to data wherever it is, but then [14:17] how do we help you standardize a little [14:18] bit so you're not doing 10 things 20 [14:20] different ways. [14:22] >> Awesome. Thank you, Sam. [14:24] >> Yeah, I think you bring up a good point. [14:25] I think, you know, we're back to [14:26] daydream again, right? It's the art of [14:28] the possible. And so, we've encouraged [14:29] our workforce to to think about all [14:31] these agents. And then on the same side, [14:33] it's like we're like, whoa, whoa, whoa. [14:34] We can't build all this. you can't have [14:36] access to this or or how are we going to [14:38] do this? And so I think there's a happy [14:39] medium we have to find, right, of how do [14:41] we, you know, you mentioned [14:42] experimentation. It's like I think [14:44] there's a whole lot of playgrounds that [14:45] are about to be turned up, right? [14:47] Because you know, Jody, you mentioned [14:48] this earlier. It's like the the only way [14:49] to do this is sometimes just to do it, [14:51] right? But we've got to do that in a [14:53] confined space and now how do we deal [14:54] with copy data management and security [14:56] and you know, red teams and blue teams [14:58] and how all this stuff's going to work. [14:59] So I think there's a lot before we get [15:01] to scaling agents that has to do with [15:03] how we even you know run and understand [15:05] and and operationalize these agents to [15:07] begin with. [15:08] >> Thank you all for being really you know [15:10] honest about the current state of even [15:12] the topic is scaling. So maybe let's [15:15] talk about some of the real challenges [15:17] and what are the scar tissues that you [15:20] are building right now which are the [15:21] muscles that we must need but maybe by [15:24] raise of hands can we see if you are [15:28] actually experimenting or building or [15:30] deploying agents even it's 1 Z 2 Z [15:34] oh that's fantastic so we will open up [15:36] for questions and I I like usually to [15:39] make it interactive and it is almost end [15:40] of the day rather than you know having [15:43] questions. So please in a few minutes [15:45] we'll open up for questions. How about [15:46] that? Okay. I think it is always good to [15:49] kind of let the audience ask you if you [15:51] all are okay with that. Right. [15:52] >> Awesome. Thank you. So Jodi let's talk [15:55] about um you know what are scars that [15:58] you are experiment know finding. Now [15:59] talk about some of the agents at you [16:02] know at Cox Automotive you are building [16:04] right now. [16:06] I think you asked me what kind of scars [16:08] we've also built. But um um I think of [16:11] our I think of our agent building in [16:14] kind of three categories. There's um the [16:16] things we do for ourselves. When I say [16:18] ourselves, I mean how we build software, [16:20] right? That job's changing. I think of [16:23] how we help the rest of our businesses. [16:25] So those are technicians in the field or [16:26] people out at our auctions who are [16:28] physically you know doing physical jobs [16:30] that we might help and build agents to [16:31] help them get their job done, do the [16:33] paperwork part of their job or um you [16:36] know aid somewhere in their workflow. Um [16:40] and then the third bucket because we're [16:41] a software company is what do we go to [16:43] market with, right? What do we offer to [16:44] our customers that's agentic based? Um, [16:48] right now we are trying to take the [16:50] lessons we learned on the first two to [16:52] see if we can make it easier for us to [16:54] do the third thing. Right. Your [16:56] customers, like you said, they trust [16:57] you. And so you're trying to make some [16:59] calls about like, yeah, I guess we'll go [17:02] to town on that platform or I will [17:04] deploy on that or I won't on that. Um, I [17:06] think probably like everyone in this [17:09] table though and everyone in this room [17:10] like we're all trying all these new [17:12] technologies. We're trying different [17:13] platforms. We're figuring out what [17:15] works. We're figuring out what doesn't [17:16] work. Um, I think we are all going to [17:20] make mistakes. We are all going to build [17:21] stuff that's too expensive or we're [17:22] going to overengineer something. And I [17:24] think that's not a scar. That's not [17:27] really a scar you should wear. You [17:28] should wear it proudly because you learn [17:31] something, you didn't do it twice. Um, [17:32] and I think that's more of the mentality [17:35] of getting through this than anything. [17:38] >> Awesome. Thank you. [17:39] >> Yeah, I think um the we're the same way. [17:43] Um but the scars that we're using for [17:45] internal processes are are driving our [17:48] decisions externally. Um and so we you [17:51] know some of the challenges we work [17:53] through is we're in a a multicloud [17:55] hybrid cloud environment. Uh and so [17:57] trying to unlock the ability for people [17:59] to use any of those things um across [18:01] that whole ecosystem wherever the data [18:03] lies and that has been incredibly [18:05] difficult and we've learned some m had [18:07] made some mistakes and sent too much [18:09] data over pipes that cost a lot of money [18:10] and had to figure out what to do to get [18:12] it back. Um but we've done that [18:15] internally so it hasn't increased our [18:17] risk exposure externally to our [18:19] customers which are really important to [18:20] us. Um, and then we can apply those [18:23] learnings in use cases that are really [18:25] close to the customer. Um, so knowledge [18:27] management is a is one that immediately [18:29] touches the customer. We did it wrong [18:31] the first time, we're rearchitecting [18:33] now, but as we do it better, we know [18:35] what the customer will consume with it. [18:37] So then we can unlock that value for [18:38] them. And so most of our scars have been [18:40] kept inside the organization, but we do [18:42] wear them proudly um so that we can then [18:44] inform how we're going to do it [18:46] externally to our customers. [18:47] >> Fantastic. [18:50] Yeah. So, um, externally in partnership [18:54] with our customers, the agents that that [18:55] we find most exciting right now are [18:57] agents that are delivering context to [19:00] our users on their enterprise data to [19:03] applications that they've already built [19:05] with LLMs. So it's this notion of a [19:08] context uh enterprise context engine [19:11] that sits in the platform that's run [19:13] with an agentic workflow that's [19:15] constantly providing context from your [19:17] enterprise data to an application driven [19:20] by an external LLM. You know going [19:22] beyond the notion of uh whether it's [19:24] prompt engineering or or rag or context [19:27] engineering by hand you know using an [19:29] agentic workflow as part of that process [19:31] is something we find super super [19:32] exciting. [19:33] >> Fantastic. you use the word context [19:35] engineering. I would love to double [19:37] click on that maybe in a few minutes. [19:39] Awesome Sam. [19:41] >> Yeah, I think uh you know you talk about [19:43] scars. I think data and the word breach [19:45] go together way more commonly than we [19:48] care about. [19:49] >> And you know we we talk to a lot of [19:51] customers too, right? And and I think [19:53] that's we've seen scars, right? From do [19:55] I turn these agents on? How much access [19:57] do they have? And I think that's the [19:59] real fear, right? Is it's a new attack [20:00] vector. It's like how are you securing [20:02] those platforms? You know, we look at [20:04] that MCP is this great server service, [20:07] whatever we want to define it as, that's [20:08] still being ratified, right? And it's [20:09] this new kind of AI platform on steroids [20:12] that allows tooling and and there's not [20:14] a standard. How do you adhere to [20:16] something, right? How do you make sure [20:17] that as fast as a GPU can communicate, [20:20] these two systems didn't just create [20:22] data or allow access to data that didn't [20:23] exist? And so I, you know, even [20:25] internally like we've worried about [20:26] these particular things, right? It's [20:27] like how much access do you give an [20:29] agent? How much freedom does it have and [20:31] where's the observability around it? [20:33] >> Absolutely. So let's unpack what do we [20:36] mean by scaling like when we talk about [20:38] like which is a even though we may not [20:40] be building at enterprise scale yet but [20:43] I think understanding will help us to [20:45] see what are the concerns areas of [20:48] concerns like when we talk about scale [20:50] whether you know whether it is the [20:51] compute whe like are you scaling at the [20:54] compute layer are you scaling at the [20:55] orchestration layer integration layer [20:57] are you scaling can we start talking [21:00] about so what do we mean by you know [21:02] scale each of you maybe starting with [21:03] them. [21:05] >> Yeah. Um, so scale I think enterprises [21:07] are at a fork in the road, right? So we [21:09] we look at unstructured data used to [21:10] live on these NAS platforms and they [21:12] weren't always the most performant, [21:13] right? So everybody's looking to say I'm [21:14] ready to refresh. How do I do this? And [21:16] what system do I pick? Right? Do I go [21:18] with a closed system, this allin-one [21:20] button? Do I go with a bunch of open [21:21] source to put this thing together? But I [21:23] think there's a futurep proof mantra [21:24] right now that a lot of leaders are in [21:26] this kind of scale paralyzed fashion [21:29] because they're trying to make a [21:31] technology choice. Where should that [21:32] data live? Where am I going to run that? [21:34] Am I going to run it on the edge in the [21:35] cloud? Am I going to compete for GPU [21:37] hours across any neo cloud that's out [21:39] there? And so I think scaling really [21:41] comes you've got to think about [21:42] foundational portions, right? To your [21:44] point, compute's gotten commodity. I [21:46] think data and storage, right, are still [21:48] this foundation thing. It's going to [21:49] have to have that to run. Agents are [21:50] going to have to make decisions off of [21:52] it. So I think some of this stuff is [21:54] shifting of what's the foundation you're [21:56] building underneath these agentic [21:57] platforms and is that particular [21:59] platform you're choosing going to scale [22:00] to meet your needs in 2026 2027 and [22:03] whatever 28 looks like. [22:05] >> Sure. [22:06] >> I think we as an industry broadly feel [22:08] comfortable at this point with managing [22:10] data at scale. I mean for for me when I [22:13] think about scaling agents or scaling [22:16] you know really any workflow in the AI [22:18] space right now I I think the main [22:20] lessons learned are around price [22:22] performance it's you know and those were [22:24] the uh were they are still the principal [22:26] challenges in enterprise data space [22:28] period so you know it's price [22:30] performance making the right decision in [22:31] terms of what types of of models to use [22:34] for a given application whether or not [22:36] you get the ROI out of an agentic [22:38] workflow that you expect to or else [22:39] you're spending more money than than [22:40] you're getting out of the application. [22:42] Um, but that all then breaks down to [22:44] price performance or price versus ROI. [22:48] >> I think I think there's a there's [22:50] actually two dimensions of scaling that [22:52] I I think about. One is the the cost [22:54] performance of scaling um and how you're [22:56] going to scale uh horizontally with more [22:59] and more and more agents within the [23:01] organization. But then I think there's a [23:03] horizontal or a vertical scaling of a [23:05] single agent throughout the [23:06] organization. And so back to our [23:08] literacy uh item um if if we can get an [23:11] agent proliferated throughout the [23:13] organization or an orchestration or a [23:15] mother agent proliferated out through [23:17] the organization and that provides value [23:19] to us, we would do that with one versus [23:21] scaling all you know getting hundreds [23:23] and hundreds of agents. And so you have [23:25] to look on those two dimensions to make [23:27] sure that you're responsibly introducing [23:29] new agents into the infrastructure so [23:31] that they're doing something meaningful [23:33] to the organization. But when you find [23:35] the ones that actually truly get ROI, [23:37] you scale and kind of squeeze that one [23:39] as much as possible to get as much [23:41] return for the organization. [23:43] >> Jordan, [23:44] >> I think I would actually echo your [23:46] answer a little bit. Watching Walmart [23:47] say like, "We had too many and now we're [23:49] going to go back and have like a [23:51] customer agent or a sales agent." [23:54] Like I really want to interview those [23:55] people because like there's a story [23:57] under there and I want to know what it [23:58] is. But to your point, I think that's [24:01] the part I think is not well understood [24:03] when I think about scaling them. Do I [24:04] need more agents or do I need one good [24:07] agent to do 10, you know, to to serve a [24:10] purpose in 10 areas in terms of [24:11] composability? Um, and my biggest [24:13] takeaway from this morning was I'm not [24:15] going to call it a mother agent. I mean, [24:16] if someone lets me, but um I thought the [24:19] Softmax guy, I was thinking like we [24:21] should have a um like a shared ethics [24:24] agent or something like or what are our [24:27] values as a company? um and that any [24:29] agentic framework, you know, runs under [24:31] that. I I don't exactly know how to [24:32] build that, but I think that's an [24:34] interesting piece of scaling agents to [24:36] say we are going to scale agents under [24:37] an ethos of things we're willing to do. [24:40] >> Thank you. I would add like from my [24:42] vantage what the way I look at it, it's [24:45] like building any software. It's like [24:47] building any software architecture. So [24:49] how we want to organize the software, [24:52] right? What is the levels tiering of you [24:56] know we need do we you know have agent [25:00] just like if I have a capability [25:01] taxonomy which might be organized by [25:04] domains or functions should we think [25:06] about that way your agents are also so [25:09] that by design I have a you know access [25:12] control what data does it have access to [25:15] what is the data product do I enable [25:17] right so kind of having method to the [25:19] badness about how we build software And [25:22] the other side is all the technical [25:24] agents like your yeah governance the [25:26] ethics the value organ those kind of and [25:30] having the you know every everything [25:32] that happens there are like three types [25:34] of agents like technical agents are [25:36] always deployed like who is observing [25:39] who is you know making sure that just [25:41] because you can doesn't mean you should [25:43] do like having that checks and balances [25:45] kind of a things so so that's how we are [25:47] approaching and I think we can I'm also [25:50] a little bit skeptical about that only [25:52] four super agent. So what does super for [25:55] super agent you know Walmart Suresh is [25:57] talking about would be nice to have a [25:59] you know double click on that [26:02] >> yeah yeah yeah so let's talk about I I [26:06] think we talked about agent now let's go [26:08] from system agent to people like so how [26:12] are you all thinking about in this [26:14] agentic enterprise what are you doing to [26:16] uplevel your people [26:20] >> and you're giving them more information [26:22] so they can do their job better Right. [26:24] Um it's I you know I said that there's a [26:27] degree of um you know an agent need to [26:29] be autonomous. Um but you know the human [26:32] in the the loop and making the human [26:34] more effective at their job is a [26:36] critical part of why we put the agent in [26:38] place right so I I think in you know 90 [26:41] plus plus percent of workflows you know [26:43] we're not taking human out of the loop [26:45] we're making the human more effective. [26:47] >> Yeah. Yeah. [26:49] I I would tweak that and say that you're [26:52] removing the failure or the errors that [26:54] the human introduces into the workflow. [26:56] >> Sure. [26:56] >> Um so like we think of it as like or I I [26:59] personally think of it as the ability to [27:01] remove all of the the repetitive [27:04] errorprone tasks that someone does um so [27:08] that they can do more value ad for the [27:09] organization. And so my goal is to put [27:12] an agent in place to remove large [27:13] portions of people's roles so that they [27:16] can do more valuable things for us and [27:17] bring more intelligence and innovation [27:19] to the organization rather than like the [27:21] standard stuff that they have [27:23] historically done [27:24] >> which is going to make you more excited [27:26] about your job too right I mean if [27:27] you're creating more value at a higher [27:28] level [27:28] >> 100%. [27:30] How about Judy? I would I couldn't agree [27:33] more, which is I think if we especially [27:35] like when we're we're building tech to [27:37] put in the hands of like someone in the [27:38] field, I'm thinking about like we we [27:41] went rode in their truck now because [27:44] what's the part of your job you hate? [27:45] That's the first thing to automate, [27:47] right? Um I know our chief product [27:49] officer who's sitting over there also [27:51] says about the people this is a whole [27:52] company transformation and that's hard [27:55] and we can't really brown bag our way [27:57] out of it. Um, so we are trying, you [28:00] know, we are working both the top and [28:02] the bottom because like I said earlier, [28:04] it's not just what we go to market with. [28:05] It's like what we do for ourselves and [28:08] getting people who don't want to use AI [28:11] to use AI to make their job more [28:12] productive. It's got to be viewed as a [28:14] helper and not as a replacer. Um, or I [28:17] think you'll see them really dig in [28:18] their heels. [28:20] Do you get are you getting a pressure [28:22] from the whether your board or the CEO [28:25] around which agents now apply what kind [28:29] of cost savings are you going to get and [28:32] and the reduction in workforce like I [28:35] have x number of employees what that [28:37] should look like and what's the you know [28:40] cost of writing business [28:42] >> you're just trying to make us touch the [28:43] hot stove aren't you [28:46] >> okay I will I will go first only because [28:48] Um, our AI accelerator has identified [28:51] something like the $20 million range of [28:53] efficiencies. So, we built 27 PC's in [28:55] the first six months of the year. It was [28:57] about making us more efficient. But I [28:58] will say we're lucky in the case that [29:00] some of that is about saying I don't [29:02] need to hire uh a a firm to do content [29:06] generation. And so, that's not your [29:08] employees, right? That's like reducing a [29:10] contract with someone. I don't need this [29:11] offshore firm to read in PDFs for us [29:13] because we could just do that ourselves. [29:15] And so, um, I won't lie. I think hitting [29:18] those targets first is so much easier. [29:21] It's so much less stressful on everyone [29:22] and it can also throw money at the [29:24] bottom line because we are still in the [29:26] business of making money. Um, and I'm [29:30] infinitely more excited about going to [29:31] market with products um that we can [29:33] charge for because that is more fun. [29:36] >> Yeah. And I think to your point, just to [29:38] kind of elaborate on it, I think it's an [29:39] opportunity to bring in specialization [29:41] into the organization that you don't [29:43] necessarily have and would have to scale [29:45] with FTEES. And so we would we would [29:48] scale that by bringing in thirdparty or [29:50] partners to do things because we don't [29:52] have the specialization. We don't have [29:54] the ability to build that specialization [29:55] within the organization. But if we can [29:57] supplement the existing um employee [30:01] population with skills that they [30:03] wouldn't have historically had, then we [30:05] can, you know, bring more of that [30:07] specialization in without having an [30:09] impact on the organization. And so it's [30:11] actually in thirdparty contracts that [30:14] will see an impact rather than to the [30:16] organization uh team itself. [30:18] >> That's wonderful. Thank you, Jim. [30:21] >> We're putting every dollar back into [30:22] innovation and I think that's what most [30:24] software companies are doing. [30:26] >> 100%. Sam, [30:27] >> yeah, I was thinking about this like [30:29] budgets now have two components, right? [30:31] I I think we've had to remind our [30:33] leadership sometimes that R&D is still [30:35] real, right? And just because it's in a [30:37] computer and in an agent doesn't mean [30:38] it's not the same as like a lab bench [30:40] that used to take three months to do [30:41] things. And so one is that we've we've [30:43] kind of budgeted to say, hey, there's [30:45] going to have to be money set aside for [30:46] us to experiment, right? The second [30:48] piece is the ROI thing is that we can't [30:50] experiment forever. At what point, how [30:52] quickly do we fail, right? Failing fast [30:53] is a big deal for us and those type of [30:55] things, too. And so, you know, I think [30:57] there's a lot that goes into the [30:58] presentation layer of what we're trying [31:00] to accomplish because I really do think [31:02] this agentic workflow is, you know, you [31:04] read about it in in, you know, CRN, pick [31:06] your magazine. It's like, well, yeah, [31:07] these agents just pop up and they do [31:09] something for you. I made the joke the [31:10] other day, I was like, my next job, I'm [31:11] I'm going to negotiate an agent, not an [31:13] admin. But like the idea is that this [31:15] digital companion is something that that [31:17] we like, right? And you see all the HR [31:19] stuff that's coming out too about how [31:21] much more interaction we're going to [31:22] have between digital human interaction. [31:24] And you know, as we get into digital [31:26] human things and and you know, looking [31:28] at at retail and entertainment, like I [31:30] think there's a new level of interaction [31:32] that we're going to have to learn to [31:33] have. And I think there's a sentiment [31:35] that comes along with that that's not [31:36] always in dollars, right? But it's about [31:38] people's peace of mind. It's about their [31:40] fear of not messing up at work, right? [31:42] And it's about the the integrity that [31:43] they have of saying, "Hey, I did my job [31:45] well today. I accomplished this today [31:47] with the help of AI. [31:49] >> Yeah. So what what trends do you are you [31:52] watching or you think innovations [31:54] talking about innovation putting dollars [31:57] back to innovation but the industry must [31:59] innovate in order for as we get into [32:02] really scaling agents right what are you [32:05] more excited about most what innovations [32:08] needs to happen you know what just tell [32:11] us something [32:15] >> I'd say small language models deployed [32:17] on the edge because they're [32:21] short and sweet. Um, I'd love to see and [32:24] I'm excited about kind of the [32:25] proliferation of more standards for [32:27] interoperability. Um, so, you know, MCP [32:30] is the buzzword of of the month. A toa [32:33] was back when during Google Next. And [32:35] so, I'd love to see more companies push [32:37] those, adopt those, and allow us to have [32:40] an open ecosystem where these things can [32:42] interact together u rather than this [32:44] kind of patchwork of connectivity. [32:49] Uh I'm super excited as we think about [32:51] price performance uh uh hybrid [32:53] deployments and hybrid deployments [32:54] becoming more effective. Um you know [32:56] these tools uh all grew up in the cloud. [33:00] Great, fantastic, but there are a lot of [33:02] companies that are going to want hybrid [33:03] as a part of their architecture from a [33:05] cost or privacy perspective. [33:07] >> Can you elaborate a little bit more? [33:09] Well, whether it's running a full stack [33:11] on prem, you know, you could have Nvidia [33:13] um Delli factory and a NIM stack, you [33:16] know, sitting next to an enterprise data [33:17] warehouse and have full stack on prem uh [33:19] as one example. Um another example is [33:22] you know you might have your data on [33:24] prem and you might move vectorized data [33:26] in the cloud and you know you're [33:27] comfortable from a privacy perspective [33:29] and data ingress egress perspective [33:31] where you're running kind of a split [33:32] architecture where you have some data on [33:34] prem and some training and and compute [33:36] done in the cloud. all those paradigms [33:38] like those paradigms didn't really exist [33:40] in an effective way to manage a few [33:42] years ago and now they're becoming [33:44] easier to manage and easier to deploy as [33:46] opposed to science projects. [33:48] >> What needs to be true in order to have [33:50] that seamlessly? [33:52] >> It's all about the management stack, [33:54] right? Um Oh, the management stack and [33:56] if you can get Nvidia to to ship you the [33:58] GPUs you need. Um but um it's all about [34:00] the management stack. Like I mean the [34:02] software exists everywhere. Um, but it's [34:04] being able to manage that software [34:05] effectively in your own data center. [34:07] That's typically where the price [34:08] performance challenge becomes you end up [34:10] on the wrong side of that because the [34:12] difficulty of managing the software. [34:14] >> No, thank you Sam. [34:16] >> Yeah, I uh, you know, we were talking [34:19] the other day and we made a joke about [34:20] we're kind of in the world of data, [34:22] we're back to clean your room, right? I [34:24] I think data sat for a long time and and [34:27] you know, kidding aside, I think the [34:29] trend of data management, right? where [34:31] we used to deal with just hey I'm going [34:33] to stuff it in the corner right it's [34:34] going to sit in SharePoint forever like [34:36] I hope to god Id never have to find [34:37] version seven of that file again is now [34:39] real right we have agents we have models [34:42] looking at our data now we're trying to [34:43] weigh the value of data and so I think [34:45] you're going to see this resurgence you [34:47] know we've always talked about [34:48] intellectual property and what that [34:49] means to our corporations but I think [34:51] all the sudden right when that data is [34:53] at the fingertips of an agent flying at [34:55] millions of flops faster than a human [34:58] can type I think we're going to have a [35:00] new trend of how do we manage that? [35:01] Where does it go? You know, you made the [35:03] comment about we we've done data at [35:04] scale. I think there's a new scale of [35:06] data. You know, you look at the stats of [35:07] like the 300 zetabytes that are going to [35:09] be created by 2027. And like I've never [35:12] used those words in my career before, [35:13] right? Like we we're talking about new [35:15] magnitudes of data of processing of [35:18] networking that we've never dealt with. [35:19] And so I think you know you guys kind of [35:21] mentioned the management side of that. I [35:23] think it's how do you get your arms [35:24] around that particular resource pool? [35:27] >> Thank you very much. questions in the [35:30] room. Any of you wants to have you know [35:32] ask questions to this brilliant panel. [35:36] >> Thank you. I'm really curious because [35:38] I'm a little hyper skeptical. So I'm [35:39] curious what um sources you guys follow. [35:43] You you represent a lot more I think [35:45] honest and transparent language than we [35:47] hear a lot of times in the industry [35:49] around scaling and and cost and [35:51] feasibility and power performance and [35:53] all of that. So, I'm curious what your [35:55] sources are and how you cultivate your [35:57] own independent thinking on this. [36:02] >> That's a great question. I I think I'll [36:03] start and and I have the easy job, [36:05] right? Because my entire job is to talk [36:06] to customers all day, right? Like I hear [36:08] these stories in every single Zoom call [36:10] I sit on, right? Even the definition of [36:12] an agent, what it is, and you start [36:14] asking probing questions. You get into [36:15] these things. You're like, "Well, you [36:17] know, your story is different than your [36:18] story, and your data looks a little [36:19] different than that data." I mean, you [36:21] know, there's a tons of podcast stuff or [36:23] whatever, but I honestly think, you [36:24] know, things like this where I get to [36:26] talk to people at the booth all day, [36:27] where we get to talk about our [36:29] challenges and pre- panel prep, those [36:31] type of things. I think humans are still [36:32] really good about talking to humans and [36:34] not this paid for advertising, who [36:37] sponsored which podcast type of stuff. [36:39] And so, you know, we talk a lot about [36:41] the community aspect of our product, [36:43] right? So, we're big into open source, [36:44] that sort of stuff. And so, but we want [36:46] that to extend further into like [36:47] personal communication because I think [36:49] that's still how we learn. [36:50] >> Do you still pay Gartner? [36:53] >> We we do. Um [36:55] I I think there's always a pay-to-play [36:57] motion in this particular industry. Um [37:01] but I think you're starting to see even [37:02] agents, right? Like I do half my [37:04] research through Perplexity or whatever, [37:06] right? as I'm trying to look at [37:07] competitive analysis between products [37:09] and I think you're going to see a new [37:10] authenticity pop out that can't be [37:12] hidden inside like market anymore, [37:14] right? [37:18] >> I mean I don't same thing. Um AI Daily [37:22] podcast is a great podcast. Um but other [37:24] than that I agree just chatting with [37:26] people. [37:28] >> Yeah. Yeah. I mean I'm going to echo [37:30] these guys. I mean clearly we all can [37:32] talk. Um, and so talking with people, my [37:35] team really hates it because I love to [37:38] do demos. I love to bring in c potential [37:41] uh partners to do demos and half the [37:43] time they tell me, "We already have that [37:45] or we already have something like that." [37:46] And I'm like, "Yeah, but they may have [37:47] something unique that we didn't even [37:49] consider." Um, and so I schedule more [37:52] demos and talk to as many partners as [37:54] possible. Um, and get the eye from my [37:58] team as a result of that. Um, same [38:00] podcast. I also like the rundown. And I [38:02] think they do a good job of little [38:03] snippets on a daily basis. So if you [38:06] miss something at 5:00 p.m. the day [38:07] before, it's in your inbox at 9:00 a.m. [38:10] so you know exactly what happened in the [38:12] last 6 hours in the AI space. [38:14] >> All right, I'm going to start listening [38:16] to that myself. Um, I challenge myself [38:19] to read one white paper a week just [38:21] because I like the math part. [38:23] >> People read those? [38:24] >> Yeah. [38:25] >> Okay. [38:25] >> Some of us. It's hard to choose one [38:28] though, actually. Um, and then I just I [38:31] do I do have several newsletters in my [38:33] inbox that come in every day, you know, [38:34] the Touring Post or other AI [38:35] newsletters. And I usually tend to piece [38:37] together from there what's true or not [38:39] true to be candid. And [38:42] and then basically it's phone a friend [38:43] and time, right? I just rely on my peers [38:45] to be like, "Hey, that doesn't seem [38:48] true, right?