Workato Rebuilds Platform to Bind AI Models to Deterministic IT Workflows

Workato 重构平台，将 AI 模型绑定至确定性 IT 工作流 AI Agent 仍受困于脆弱的边缘聊天机器人，原因在于随机性语言模型缺乏企业自主运营所需的确定性执行技能与实时运营信号。 Adam Seligman 解释了为何现代 AI Agent 仍停留在浅层的生产力边缘，以及释放真正的运营自主性为何需要深度的企业级编排。他概述了企业上下文的三层模型：从静态存储数据，演进到流程可见性，再到实时计算的业务信号。为确保安全性与可靠性，Workato 将随机性 LLM 推理与确定性、可复用的企业技能相结合；Seligman 结合 Workato Go 等集群管理工具、Lead Rocket 外拓销售 Agent 以及 Serverless MCP 实现对该方案进行了阐述。 尽管基础设施投入巨大且大模型快速迭代，当前多数企业 AI 仍停留在边缘生产力工具和脆弱聊天机器人阶段，缺乏真正的企业级自主执行力。 (3:00) 实现真正的企业级 Agent 必须跨越三个上下文层级：从静态数据库查询和端到端流程可见性，升级至由实时业务信号驱动系统级问题修复。 (5:00) 企业需要将概率性的大语言模型与确定性执行相结合，通过将多步骤集成工作流转化为安全、可复用的企业技能（Skills），并围绕明确的 KPI 指标构建核心 Agent。 (8:14) 通过 Workato Go 等集中式管理平台与托管式 MCP 服务，企业能够高效监控自主 Agent 集群运行状态，并将复杂业务流程封装为结构化的模型调用工具。 (11:58)

Workato CTO Adam Seligman argues that despite massive capital investments, enterprise AI remains confined to peripheral productivity tasks because language models lack operational context and deterministic reliability. To achieve true enterprise agency, organizations must bridge stochastic model reasoning with codified, multi-system integration skills and live operational signals. The AI Paradox and Context Limitations (3:00) Despite an estimated quarter-trillion dollars in infrastructure capital expenditures and rapid model advancements, Seligman observes that enterprise AI remains largely confined to peripheral tasks like email autocompletion and basic customer support chatbots. In these edge deployments, users are often forced to manually guide fragile tools that present incorrect outputs with misplaced confidence. Furthermore, organizations face strict governance hurdles that prevent unconstrained agents from accessing sensitive systems like payroll. Seligman attributes this edge stagnation to a fundamental lack of operational context and trust. He divides enterprise visibility into three tiers: static database records (Level 1), end-to-end process visibility (Level 2), and live business signals (Level 3). While Level 1 merely confirms an order exists and Level 2 tracks carrier handoffs, Level 3 uses live operational metrics—such as real-time delivery rates or tariff disruptions affecting dozens of shipments—to prompt systemic business remediation rather than resolving isolated support tickets. Pairing Stochastic Reasoning with Deterministic Execution (8:14) Enterprise IT environments require deterministic, predictable execution, whereas large language models operate on probabilistic, statistical logic. Organizations cannot risk hallucinated or intermittent execution when provisioning employee credentials, issuing invoices, verifying credit, or processing payments. To bridge this gap, Workato codifies multi-step integration workflows into reusable, deterministic enterprise "Skills." Under this architecture, which underpins the Workato One platform, an LLM provides non-deterministic cognitive reasoning while codified skills enforce transactional integrity, state compensation, and security. Seligman specifies that a viable "Core Agent" must be anchored to explicit business KPIs and ingest live contextual signals across enterprise applications, translating high-level reasoning into verified operational tasks across more than 12,000 supported applications. Autonomous Fleet Management in Practice (11:58) Operating enterprise agents at scale demands governance frameworks where human operators supervise autonomous actions. Workato demonstrated this via Workato Go, an executive interface that tracks agent fleets against defined business targets, surfaces real-time operating metrics, and alerts humans when exceptions require intervention. As an operational case study, Seligman highlighted Lead Rocket, an autonomous outbound sales agent. The tool monitors business and market signals to identify prospect accounts, executes personalized email campaigns, routes responses to human sales representatives, and runs autonomous experiments—such as detecting when a client champion changes employers to initiate targeted outreach at the new organization. Deep Enterprise Orchestration Beyond Raw Protocols (13:48) While Anthropic's Model Context Protocol (MCP) establishes an open standard for LLMs to communicate with external tools, Seligman notes that major SaaS vendors remain protective of their proprietary data silos. Additionally, exposing raw, comprehensive API schemas directly into a model's context window can overwhelm its reasoning capacity. To address this, Workato provides serverless MCP servers that bundle complex, multi-application integration recipes into focused business endpoints. However, Seligman cautions that protocol-level access represents only the surface layer of integration. Production deployments require deep, out-of-band orchestration capabilities that protocols alone do not provide, including user-level identity management, transactional retry mechanisms, state compensation, and resilient exception handling to prevent workflow corruption. Seligman concludes that scaling enterprise AI requires adopting purpose-built orchestration platforms rather than building custom infrastructure from scratch. However, full enterprise autonomy remains contingent on resolving data access barriers across protective SaaS vendors and enforcing strict governance across autonomous fleets.

Workato首席技术官Adam Seligman指出，尽管全球在人工智能基础设施上投入了数千亿美元，但当前AI代理仍因缺乏企业上下文与确定性执行力而局限于浅层生产力边缘；企业若要实现核心业务的真正智能化，必须将大语言模型的概率推理与具备业务上下文的确定性工作流深度编排。 AI落地的边缘悖论与安全瓶颈 (3:00) Seligman表示，尽管行业正经历前所未有的技术创新（如GPT-5、Claude 3.5 Sonnet、Claude Opus以及AWS新代理工具的推出）并伴随着约2500亿美元的基础设施资本支出，旧金山等技术中心每周举办约60场AI研讨会，但绝大多数AI应用仍被困在“生产力边缘”。当前的落地形态多局限于邮件自动补全、基础回复草拟以及死板的客服意图识别。用户面对的往往是脆弱的聊天机器人，不仅频繁给出错误但看似自信的回答，还需要人类像“推着蹒跚学步的幼儿上楼梯”一样持续监督和修正。 与此同时，企业严苛的安全合规底线构成了根本性阻碍。大型组织无法直接将未受约束的代理引入薪酬计算、核心财务数据库等敏感业务场景。由于缺乏健全的治理框架，企业不敢赋予AI真正的自主行动权，导致巨额资本与前沿算法的投入与企业核心业务运作之间出现了显著脱节。 迈向动态业务信号驱动的第三层上下文 (5:00) Seligman分析认为，AI代理无法进入企业核心的根本症结在于缺乏业务运行上下文。他将企业理解客户业务的能力划分为三个层级：第一层依赖数据库中的静态存储数据（Data at Rest），系统仅能反馈孤立状态（例如告知客户“订单已在洛杉矶”），无法判断包裹是否停滞；第二层具备端到端流程可见性，系统能识别流程堵点（例如检测到货物卡在洛杉矶港，并自动通过备用承运商补发替代品）；第三层则是基于实时业务信号的系统性治理与修复。 在第三层架构中，单次客户咨询会被视作覆盖全企业的触发信号。例如，系统在响应个案时能联动发现洛杉矶共有67笔订单因关税政策调整受阻，进而直接触发全局业务流程调整，而非机械处理单个工单。Seligman指出，要支撑这种高阶自动化，代理不能依赖数据仓库中滞后的事后分析，而必须实时计算并摄取动态业务信号（如82%的准时交付率、70%的收入风险指标或实时客户满意度评分），从而具备类似资深运营主管的业务直觉。 确定性技能与随机大模型的工程协同 (8:14) 大语言模型（LLM）本质上是基于概率和统计的随机系统，而企业IT架构的核心诉求则是执行的绝对确定性。在员工账号开通、发票开具、信用审核或支付结算等关键任务中，企业无法容忍间歇性失效或不可控的试错成本。Seligman以人类婴儿学步与拿水杯为例，指出人类是通过反复训练将笨拙的动作转化为可靠的肌肉记忆；Workato则通过将验证成熟的多步骤集成“配方”沉淀为可复用的“技能”（Skills），为随机的AI推理层提供具备事务性、安全性、可解释性与容错能力的确定性执行载体。 这种架构催生了所谓“核心代理”（Core Agent）。每个核心代理都绑定明确的业务KPI目标与任务范围，而非依赖宽泛的对话提示词；代理结合端到端上下文并调用技能层在底层系统中安全执行更改。为此，Workato将其平台重构为Workato One，将原生连接超过1.2万个企业应用、本地数据库、EDI与RPA的既有集成网络，与新兴的自主代理编排体系全面打通。 代理集群治理与深层编排协议的局限 (11:58) 针对企业内部多个自主代理的协同运作，Seligman展示了集中式高管管理看板Workato Go。在Workato内部应用中，自主销售开发代理Lead Rocket能自主监控市场信号以挖掘商机、执行个性化外发邮件、向人类业务代表（SDR）分发回复，甚至在监测到既有客户的关键倡导者跳槽至新公司时，自主针对新组织发起定制化触达实验。当遇到超出自主权限的异常时，代理会主动向人类团队发起升级请求，形成协同运行的数字机队。 在接口标准化层面，Seligman探讨了Anthropic的Model Context Protocol（MCP）。他指出，主流SaaS厂商倾向于将数据封装在自身壁垒内，乐于调用外部MCP工具却极少主动开放底层数据；同时，直接将庞大且扁平的原生API模式（如整套ERP或Asana接口）灌入LLM上下文会迅速耗尽推理资源。为此，Workato推出无服务器托管MCP服务，将复杂的多应用流程抽象为针对具体业务任务的端点。Seligman强调，MCP仅解决了最外层的接口暴露，真实的生产部署依然高度依赖分布式重试逻辑、事务边界控制、状态补偿以及防范流程级联崩溃的深层编排能力。 Seligman总结认为，企业从AI概念验证走向规模化落地的过程中，自建底层编排与连接架构往往效率低下，依托成熟的企业级平台是控制风险和加速转型的关键路径。Workato计划在其WOW大会上正式推出系列企业代理产品；然而，如何在主流SaaS供应商筑起的数据孤岛壁垒下持续获取深层流程权限，并在高并发与复杂业务冲突下确保代理集群的绝对稳定，依然是未来企业代理化演进中待解的技术与生态课题。

## Background 本内容由频道 Ai4 于 2025 年 8 月 24 日在 YouTube 平台发布。视频记录了 Workato 首席技术官 Adam Seligman 的主题演讲，主要探讨在 AI 热潮中，技术管理者如何通过企业级编排突破 Agent 落地瓶颈，重塑技术栈与工作方式，构建真正的 Agentic Enterprise。 ## TL;DR Adam Seligman 指出当前 AI Agent 大多被困在企业边缘做浅层交互，根源在于缺乏深度的企业上下文和实质行动代理权。为了让不可预测的大模型在严谨的企业 IT 环境中发挥价值，Workato 提出了以业务 KPI 为核心、结合确定性企业级技能（Skills）的智能体架构。通过最新推出的 Workato One、Workato Go 智能体舰队管理平台以及托管式 MCP 服务，企业能够将散落在异构系统中的业务流程安全封装并赋予 Agent，从而实现从边缘玩具到生产级核心生产力的跨越。 ## Atmosphere & style 演讲风格务实、犀利且充满现场感染力，Adam 凭借在亚马逊等企业的深厚工程背景，对当前行业盲目的 AI 炒作进行了冷峻而精准的剖析。他运用生动形象的比喻（如“推着刚学走路的幼儿上楼梯”与“幼儿学喝水”）将复杂的企业架构理论讲得深入浅出，台下互动热烈。整场分享节奏明快、直击技术痛点，既展现了资深技术掌舵者的严谨风范，又传递出对智能体赋能企业变革的充沛信心。 ## Key takeaways - [02:13](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=133s) 业务流程而非单一应用程序，才是决定企业在市场竞争中建立差异化壁垒的核心竞争力。 - [03:00](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=180s) 当前市面上的 AI Agent 大多停留在企业边缘，充当简单的补全助手或生硬的客服路由，缺乏真正的自主性与代理权。 - [04:00](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=240s) Agent 无法深入企业核心业务的两大根本瓶颈在于：缺乏全景企业上下文（Enterprise Context）以及缺乏企业对其执行核心任务的实质代理信任（Agency）。 - [05:07](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=307s) 企业上下文由浅入深分为三个层次：Level 1 静态存储数据、Level 2 端到端业务流程感知、Level 3 实时业务健康信号与结构性缺陷捕捉。 - [08:14](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=494s) 企业级 IT 无法容忍大语言模型基于概率的随机性（Stochastic），必须将成熟业务规则固化为百分之百确定且具备事务性的企业技能（Skills）。 - [09:35](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=575s) 核心智能体（Core Agent）应采用三元架构：以核心业务 KPI 为牵引，左侧注入实时企业上下文，右侧调用安全可控的企业级确定性技能库。 - [11:18](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=678s) Workato 正式推出 Workato One 底座与 Workato Go 管理平台，支持企业统一监控并协同调度自主运行的“智能体舰队”。 - [13:48](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=828s) 在 MCP 协议生态中，Workato 提供中立的完全托管 MCP 服务器，将底层业务配方（Recipes）按任务封装，避免直接向模型暴露裸露 API 导致上下文超载。 - [15:21](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=921s) 智能体编排绝非简单的 API 串联或低代码画布连线，规模化生产落地必须依靠底层的带外安全鉴权、分布式事务一致性与高弹性重试机制。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=4s) Workato 与企业级编排的核心价值 Adam Seligman 登场并向在场观众提问有多少人了解或正在使用 Workato，现场反响热烈。作为企业集成与编排（Integration and Orchestration）领域的行业领跑者，Workato 历经十年的深耕，已在金融服务、软件、电信、零售、高科技、医疗健康和制药等各行各业建立了深厚的客户基础。Adam 提到，展示这些知名客户并非单纯为了宣扬声势，而是为了强调“业务流程”（Processes）在这些大型企业中的支柱地位。 这些大型企业并非随意摸索业务规则的初创公司，而是对推动“卓越运营”（Operational Excellence）有着严苛要求的成熟组织。结合自己在亚马逊（Amazon）工作三年半的切身经历，Adam 指出顶尖企业对运营卓越性的把控极其严谨，这也倒逼 Workato 不断打磨平台能力，以满足客户对系统高可靠性与流程规范性的极致要求。 Workato 的创立初衷非常纯粹：企业的日常运转依赖于极其庞杂的技术堆栈。企业内部通常同时运行着 CRM 系统、HCM 人力资源系统、业务数据库、运营数据库，而法务部门可能还在用 Box 和 DocuSign 协同处理合同修订（Redlines）。为了解决这种碎片化难题，Workato 构建了一个连接超过 12,000 款应用程序的企业级编排平台。无论是业务人员还是专业开发者，都能借此打通全套业务流程，驱动企业高效、有序地运转。Adam 强调，真正的“编排”绝不仅仅是在界面上拖拽几个流程步骤，而是深入企业底层，支撑起运转核心业务的关键流程。 - **10 年行业积淀**：跨越金融、电信、零售、医疗及高科技等主流行业。 - **卓越运营基因**：依托在 AWS 等顶尖企业的运营经验，专注提供高可靠的企业级服务。 - **12,000+ 应用连接器**：为业务人员和开发者提供跨 CRM、HCM、数据库与协同软件的端到端编排能力。 --- ### [02:13](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=133s) 业务流程才是企业的核心竞争力 业务流程是企业脱颖而出的本质所在，它直接决定了企业在市场中的竞争壁垒、客户服务水平、新产品推向市场的创新速度，以及员工的内部协同体验。Adam 强调，这些差异化优势是由“跨越应用程序的业务流程”带来的，而非孤立的 App 本身；流程天然贯穿于多个系统之间，并且深度包含人类员工的参与和决策。 在实际运营中，即便管理者能够查看任意单一系统中的静态存储数据（Data at rest），或者利用数据分析工具进行事后复盘（Analytics），也无法真正掌握业务流程当前的实时运行状态与健康度。流程跨系统的动态特性，使得单纯依赖孤立系统的数据监控难以反映全局业务的真实脉搏。 --- ### [03:00](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=180s) 走出边缘：当前 AI Agent 的落地瓶颈与行业虚火 谈及当前火热的 AI 浪潮，Adam 直言不讳地指出了一个关键痛点：目前市面上的各类 Agent 几乎都被困在企业的“边缘”（Stuck at the edge）。它们主要充当生产力工具中的自动补全助手，或是客服系统里生硬猜测用户意图的路由分流器。这些工具经常给出完全错误的判断，业务范围被严重窄化，却往往表现出“盲目的高度自信”。 ```mermaid flowchart TD accTitle: Agent 发展现状与困境 subgraph Bottlenecks["当前 Agent 局限 (Stuck at the Edge)"] A["边缘任务与局部补全<br/>(邮件补全 / 简单客服分流)"] B["缺乏自主性<br/>(如推幼儿上楼梯般依赖手动 Prompt)"] C["安全与权限隔离缺失<br/>(无法连接薪酬系统等核心数据)"] end subgraph Causes["核心症结 (Two Root Causes)"] D["缺乏企业级上下文<br/>(Lack of Enterprise Context)"] E["缺乏实质行动代理权<br/>(Lack of Agency / Trust)"] end Bottlenecks --> Causes ``` 这种缺乏真正自主性的交互模式被 Adam 形象地比喻为“推着刚学会走路的幼儿上楼梯”（pushing along like a toddler up the stairs）——用户必须在每一步都小心翼翼地敲击键盘引导，根本谈不上自动化。此外，在企业安全层面，如果没有严密的权限管控与基础设施，企业根本不敢让 Agent 接触核心资产，"千万别把薪酬数据（Payroll）接进去"，因为没有人能预测 Agent 会做出什么举动。 与此同时，整个行业正经历着前所未有的狂热投入：GPT-5 蓄势待发、Anthropic 发布 Claude 3.5 / 4 等系列模型、AWS 推出全新的 Agent 技术，以及 MCP（Model Context Protocol）和 A2A 等互操作标准相继涌现。仅在旧金山，每周就有大约 60 场 AI 研讨会；全行业累积的资本支出（CapEx）已高达约 2500 亿美元（Quarter trillion dollars）。然而，巨额的资金与技术创新依然无法掩盖 Agent 停留在边缘的尴尬现状。Adam 总结了背后的两大根本症结： 1. **缺乏企业上下文（Enterprise Context）**：没有为 Agent 提供充足的业务洞察和业务数据，如同给它们戴上了枷锁； 2. **缺乏实质代理权（Agency）**：企业缺乏对 Agent 的信任，不敢把真正重要的任务直接放权给它们执行。 --- ### [05:07](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=307s) 企业上下文的三层进阶：从静态数据到实时业务信号 为了具象化理想中的 Agent 形态，Adam 以典型的电商物流客服场景为例，展示了从传统客服到核心企业级 Agent 的三层能力进阶。当客户询问“原本周一就该送达的订单为什么还没到”时，不同阶段的技术响应截然不同： ```mermaid flowchart LR accTitle: 企业上下文的三层演进 L1["Level 1: 静态数据<br/>(Data at Rest)<br/>'订单在洛杉矶在途'"] --> L2["Level 2: 端到端流程<br/>(End-to-End Process)<br/>'滞留洛杉矶港，主动换路补发'"] L2 --> L3["Level 3: 实时业务信号<br/>(Business Signals)<br/>'关税异动致 67 单受阻，触发流程重构'"] ``` 在第一层（基础数据层），系统只能查询静态数据，机械回复“订单在洛杉矶运输中”，这对解决用户焦虑毫无实质帮助；到了第二层（流程感知层），系统深入到了订单创建、仓库拣货、单号生成到承运商交接的端到端全流程，发现货物卡在洛杉矶港口，从而能够主动为客户补发一件并更换运输渠道；而企业真正需要的则是第三层（业务信号层）——当该事件触发全局检查后，企业发现因为关税政策变动导致共有 67 个订单全部滞留在港口，这不再是单一的客户服务问题，而是一个必须立即修复的结构性业务流程缺陷。 Adam 解释道，诸如“82% 准时交付率”、“70% 收入风险”或“78 分客户满意度（CSAT）”等指标，并不是 CRM 数据库中现成的静态字段，而是关于企业运营健康状况的实时动态计算值。大多数企业以往只能在数据仓库中进行事后分析，但如果能将这种实时的“业务信号”注入给 AI Agent，就如同在洛杉矶仓库里安排了一位经验丰富、能主动发现阻塞并协调全盘解决方案的超级员工一样，这才是 Agent 赋能企业的真正价值所在。 - **Level 1 静态数据 (Data at Rest)**：跨应用查询订单、付款与单号状态。 - **Level 2 端到端流程 (End-to-End Processes)**：精确感知事件在业务链路中所处的节点与阻塞原因。 - **Level 3 实时业务信号 (Business Signals)**：动态计算业务健康度，从孤立客服事件中提炼出结构性流程优化点。 --- ### [08:14](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=494s) 应对随机性：将 Recipe 沉淀为确定性的企业 Skill 推进 Agent 落地核心业务的另一大障碍在于 AI 本身的“随机性”（Stochastic / Unpredictable）。AI 基于概率模型运行，表现为“有时奏效、有时勉强能用、有时彻底崩溃”。然而，企业 IT 架构和 CIO/CTO 们的底线是系统的高可用与确定性，“偶尔能用”（working sometimes）对于严谨的企业级 IT 环境来说是完全无法接受的。 为了化解这一矛盾，Adam 举了人类动作学习的生动案例：成年人拿起水杯喝水几乎无需思索，动作精准且稳定；但幼儿在学喝水时手脚不稳、频繁洒水，必须时刻提防桌上的“危险区域”。人类通过反复训练将动作固化为可预测、可重复的稳定技能，企业流程同样需要这种确定性。 Workato 将平台原有的集成配方（Recipes）进一步升级为 AI 可直接调用的“技能”（Skills）。例如，企业可以将“如何开具发票”的标准业务规则一次性编排并固化为一项 Skill。这样一来，不仅人类员工和后台自动化流程可以使用该标准逻辑，AI Agent 也能够以百分之百确定、可复用且受控的方式执行发票开具，成为企业构建高可信 Agentic 系统的坚实基石。 ### [09:35](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=575s) 构建核心智能体：以 KPI 为中心，融合确定性技能与大模型 在企业日常运营中，存在着数以千计必须确保 100% 可靠的核心技能，例如为新入职员工开通系统权限、生成发票、追踪付款记录以及核验客户信用等。这些业务环节容不得半点差错。如果单纯依赖大语言模型（LLM）去猜测这些业务逻辑，输出必然呈现随机性（stochastic），在很多情况下都会崩溃失效。企业级的任务执行必须满足四项基本要求：安全、具备事务一致性（transactional）、高弹性容错（resilient）以及可解释性（explainable）。 为此，Workato 提出了构建企业“核心智能体”（Core Agent）的方法论：智能体的设计必须始终以业务关键绩效指标（KPI）为核心。正如企业招聘一名优秀员工时首先要明确其使命、服务对象与岗位职责一样，构建智能体也必须先回答“它存在的使命是帮公司提升哪一项核心指标”。架构上，核心智能体形成了一个三元结构：中间是明确的业务 KPI，左侧是它所必需的企业级全景上下文（Enterprise Context），右侧则是用于执行具体操作的企业级技能库（Enterprise Skills）。 ```mermaid flowchart LR accTitle: 核心智能体三元驱动架构 subgraph ContextSide["输入与认知"] CTX["企业上下文<br/>(实时数据 / 业务背景)"] end subgraph CoreAgent["智能体核心"] KPI["核心业务 KPI<br/>(使命与职责)"] LLM["大模型推理大脑<br/>(决策与规划)"] KPI <--> LLM end subgraph ActionSide["确定性执行"] SKILLS["企业级技能库<br/>(安全 / 事务性 / 弹性)"] end CTX --> LLM LLM --> SKILLS ``` 将大语言模型的思考认知大脑、清晰的任务定义、实时上下文以及改变业务状态的行动能力有机结合，才能真正迈入 KPI 驱动的智能体新时代。当前整个行业在 LLM 框架上的尝试大多侧重于提示词和模型侧，对于如何构建右侧那些“确定性、事务性的企业技能”，业界其实缺乏沉淀，这些深水区工程能力“根本不是你在 Udemy 上看几门网课就能掌握的”。 --- ### [11:18](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=678s) Workato One 与 Workato Go：管理企业级智能体舰队 为了连接不可预测的随机性（stochastic）与严苛的确定性（predictable），Workato 历经十年沉淀，彻底重构了底层平台，正式推出 Workato One。该平台具备全栈集成连接能力：不仅能连接超过 12,000 款商业应用软件，还覆盖了各类数据库、本地部署（on-premise）系统，同时内置对 EDI 和 RPA 的全面支持。在这一极其完备的企业底座之上，Workato 叠加了智能体编排层（Agentic Orchestration），让企业能够无缝构建驱动业务增长的核心智能体。 紧随其后，演讲现场全球首发了下一代管理平台 Workato Go 的预览（此前仅在一个月前发布了初始版本，下周将在拉斯维加斯年度大会上正式发布）。Workato Go 采用以仪表盘为中心的方式，用于统一管理部署在整个企业内部的“智能体舰队”。以首席执行官 Vijay 的个人控制台为例，界面左侧显示的绝非传统的静态统计图表，而是各个活跃智能体实时自主同步的状态报告。 以内部运行的 **Lead Rocket** 智能体为例，它是一个能够自主运作的外拨潜在客户挖掘（Outbound Lead Gen）智能体： - **目标追踪与信号捕捉**：自主锚定拓客 KPI，全网捕捉潜在商机信号。 - **自动化外联与线索路由**：识别联系时机后，自主排期并向潜在客户发送定制邮件；一旦收到客户回信，立即智能分发给 SDR（销售发展代表）团队跟进。 - **自主设计 A/B 实验**：它会主动构思并测试新的外发邮件策略与商业信号。例如，当系统监测到某个曾喜爱 Workato 的老客户离职并入职新公司时，Lead Rocket 会将此识别为极高转化率的破冰机会，并自主围绕该场景设计实验方案。 这种智能体不再是被动执行单次问答的机器人，而是能够在既定 KPI 框架下自主推理、调动上下文并持续自我优化的数字员工，只有在真正遇到阻碍时才会主动“举手”向人类请求协助。 --- ### [13:48](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=828s) 中立的 MCP 服务与企业级流程编排 针对业内炙手可热的 MCP（Model Context Protocol，模型上下文协议），Workato 始终保持像“瑞士”一般的中立连接者角色。MCP 虽为大语言模型发现和调用 API 提供了一层更直观的协议封装，但它在当前落地中存在严重的生态壁垒：许多大型 SaaS 厂商表面上宣称拥抱 MCP，实际上却非常排他——他们乐意读取外界的 MCP，却极力抗拒将自身应用系统与数据库内部的业务数据通过 MCP 接口对外暴露。 ```mermaid flowchart TD accTitle: Workato 托管式 MCP 架构 subgraph Apps["企业异构系统资产"] A1["SaaS 应用 (如 Asana)"] A2["企业数据库 / ERP"] A3["自研系统 / APIs"] end subgraph WorkatoPlatform["Workato 无服务器托管 MCP"] REC["业务配方 (Recipes) 编排"] SEC["内置认证、安全与事务重试"] EP["按业务流程聚合的定制 MCP 端点<br/>(例如：订单生命周期管理)"] REC --> SEC --> EP end subgraph Consumers["智能体消费端"] C1["Claude / ChatGPT"] C2["企业私有 Agent 项目"] end Apps --> REC EP --> Consumers ``` 针对这一行业断层，Workato 推出了完全托管、无服务器（Serverless）的企业级安全 MCP 服务器： - **定制化业务端点，拒绝裸露 API**：传统的做法是将某个应用（如 Asana）上成百上千个底层扁平 API 粗暴打包成 MCP，这不仅无法帮助智能体完成任务，反而会导致模型上下文严重超载。Workato 支持企业把底层集成配方（Recipe）按照具体业务任务（如“订单生命周期处理”）进行封装，仅向智能体暴露该流程核心的操作谓词与接口。 - **即插即用**：开发人员可以针对任何连接器快速构建定制的 MCP 端点，供 Claude.ai、ChatGPT 或企业内部自研的各类智能体项目稳定调用。 --- ### [15:21](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=921s) 从玩具到生产环境：跨越扩展落地的鸿沟与尾声 回顾过去一年业内对“智能体编排（Orchestration）”的讨论，很多技术人员往往产生一种轻量化的误解，以为编排无非就是“拿 MCP 调通几个 API，再在画布上拖拽几个流程图小方块（doodly box thing）”，就能大功告成。然而在企业级生产环境中，原型开发与规模化部署之间隔着巨大鸿沟： - **企业级安全边界**：MCP 协议本身并没有完善定义跨系统安全鉴权与访问控制，大量安全能力完全是在协议带外（out of band）需要平台去硬核支撑的。 - **事务一致性与重试机制**：底层的分布式事务、失败重试（retries）、断点续跑等高可用机制，绝非那些玩具式的浅层工作流工具能够应付。 在从概念验证（PoC）走向大规模投产的历程中，企业必然会遭遇无数此类极具深度的工程暗坑。与其耗费数年时间去自研重新造轮子，更理智的抉择是直接选用一个历经十年考验、真正解决底层棘手难题的企业级集成平台。 演讲在热烈的掌声中进入尾声。Seligman 热情邀请在场的数千家客户与参会者参加下周同样在拉斯维加斯举办的 Workato WOW 年度峰会，现场见证更多企业级智能体舰队的真实落地场景，并幽默地表示：“如果大家在 LinkedIn 上私信我，我手里很可能还留有峰会的门票折扣码。” ## Notable quotes > [00:04](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=4s) 真正的“编排”绝不仅仅是在界面上拖拽几个流程步骤，而是深入企业底层，支撑起运转核心业务的关键流程。 — Adam Seligman > [02:13](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=133s) 业务流程天然贯穿于多个系统之间，并且深度包含人类员工的参与和决策。 — Adam Seligman > [03:00](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=180s) 目前市面上的各类 Agent 几乎都被困在企业的“边缘”。它们主要充当生产力工具中的自动补全助手，或是客服系统里生硬猜测用户意图的路由分流器。 — Adam Seligman > [03:00](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=180s) 千万别把薪酬数据（Payroll）接进去，因为没有人能预测 Agent 会做出什么举动。 — Adam Seligman > [05:07](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=307s) 大多数企业以往只能在数据仓库中进行事后分析，但如果能将这种实时的“业务信号”注入给 AI Agent，就如同在洛杉矶仓库里安排了一位经验丰富、能主动发现阻塞并协调全盘解决方案的超级员工一样，这才是 Agent 赋能企业的真正价值所在。 — Adam Seligman > [08:14](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=494s) “偶尔能用”（working sometimes）对于严谨的企业级 IT 环境来说是完全无法接受的。 — Adam Seligman > [09:35](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=575s) 企业级的任务执行必须满足四项基本要求：安全、具备事务一致性（transactional）、高弹性容错（resilient）以及可解释性（explainable）。 — Adam Seligman > [09:35](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=575s) 对于如何构建右侧那些“确定性、事务性的企业技能”，业界其实缺乏沉淀，这些深水区工程能力“根本不是你在 Udemy 上看几门网课就能掌握的”。 — Adam Seligman > [15:21](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=921s) 很多技术人员往往产生一种轻量化的误解，以为编排无非就是“拿 MCP 调通几个 API，再在画布上拖拽几个流程图小方块”，就能大功告成。 — Adam Seligman > [15:21](https://www.youtube.com/watch?v=O0UsASDE5Ww&t=921s) 如果大家在 LinkedIn 上私信我，我手里很可能还留有峰会的门票折扣码。 — Adam Seligman ## Mentioned Adam Seligman, Workato, Amazon, AWS, Box, DocuSign, GPT-5, Anthropic, Claude 3.5, Claude 4, MCP (Model Context Protocol), A2A, Workato One, Workato Go, Vijay, Lead Rocket, Asana, ChatGPT, Claude.ai, Udemy, Workato WOW ## Video description Adam Seligman, CTO, Workato Abstracts Amid the AI hype, technology leaders face a critical challenge: achieving real, agent-powered transformation. Join Workato CTO Adam Seligman as he shares Workato’s proven blueprint for building an agentic enterprise—developed in partnership with customers and pioneered by Workato internally as our own first customer. In this session, Workato will reveal key lessons and actionable best practices, including: Why orchestration is essential for unlocking value with AI agents How forward-thinking leaders are reimagining their tech stacks for the agentic era Why agentic transformation is as much about transforming how people work as it is about transforming technology Discover how to move beyond the hype and drive meaningful change with AI agents in your organization. [Source](https://www.youtube.com/watch?v=O0UsASDE5Ww)

[0:04] Adam Seligman: 真正的“编排”绝不仅仅是在界面上拖拽几个流程步骤，而是深入企业底层，支撑起运转核心业务的关键流程。 [2:13] Adam Seligman: 业务流程天然贯穿于多个系统之间，并且深度包含人类员工的参与和决策。 [3:00] Adam Seligman: 目前市面上的各类 Agent 几乎都被困在企业的“边缘”。它们主要充当生产力工具中的自动补全助手，或是客服系统里生硬猜测用户意图的路由分流器。 [3:00] Adam Seligman: 千万别把薪酬数据（Payroll）接进去，因为没有人能预测 Agent 会做出什么举动。 [5:07] Adam Seligman: 大多数企业以往只能在数据仓库中进行事后分析，但如果能将这种实时的“业务信号”注入给 AI Agent，就如同在洛杉矶仓库里安排了一位经验丰富、能主动发现阻塞并协调全盘解决方案的超级员工一样，这才是 Agent 赋能企业的真正价值所在。 [8:14] Adam Seligman: “偶尔能用”（working sometimes）对于严谨的企业级 IT 环境来说是完全无法接受的。 [9:35] Adam Seligman: 企业级的任务执行必须满足四项基本要求：安全、具备事务一致性（transactional）、高弹性容错（resilient）以及可解释性（explainable）。 [9:35] Adam Seligman: 对于如何构建右侧那些“确定性、事务性的企业技能”，业界其实缺乏沉淀，这些深水区工程能力“根本不是你在 Udemy 上看几门网课就能掌握的”。 [15:21] Adam Seligman: 很多技术人员往往产生一种轻量化的误解，以为编排无非就是“拿 MCP 调通几个 API，再在画布上拖拽几个流程图小方块”，就能大功告成。 [15:21] Adam Seligman: 如果大家在 LinkedIn 上私信我，我手里很可能还留有峰会的门票折扣码。

[0:04] Adam Seligman: Processes represent cross-functional, human-in-the-loop operational differentiators rather than isolated SaaS applications. [0:04] Adam Seligman: Static data at rest fails to capture the dynamic, in-flight state of critical business processes. [3:00] Adam Seligman: pushing a toddler up the stairs [5:00] Adam Seligman: Your order is on the way; it's in LA. [5:00] Adam Seligman: Your order is stuck in the Port of LA, so we have dispatched a backup pair via an alternative carrier. [8:14] Adam Seligman: Enterprise IT requires predictability and deterministic execution, whereas Large Language Models are inherently stochastic and statistical. [8:14] Adam Seligman: The cognitive reasoning of LLMs directs action, while codified Workato skills enforce strict, reliable business rules. [8:14] Adam Seligman: Every core agent is assigned an explicit business mission and measurable KPI targets rather than vague conversational prompts. [13:48] Adam Seligman: wrapping an API with an MCP endpoint is only the outermost layer of orchestration. [15:59] Adam Seligman: Rather than reinventing custom architecture to solve scaling hurdles, enterprises should rely on established platforms designed to handle complex orchestration and agent governance.

[0:04] I'm Adam Seligman. I work at a company [0:05] called Wcado. Does anybody know Wcato or [0:08] use Ricardo? Yeah, a lot a lot of hands. [0:10] Cool. So, we're a pretty well-known [0:12] integration and orchestration company, [0:14] one of the leaders in that category. And [0:16] I wanted to share a couple lessons that [0:18] I've had working at Marcato in my time [0:20] at other companies like AWS, learning [0:22] about how to get uh Agentic Systems [0:24] working at the core of your enterprise. [0:29] uh work Kado has a pretty amazing [0:31] customer base. It's been a a 10-year [0:34] journey to build incredible [0:35] relationships in pretty much every [0:37] industry from uh financial services, [0:40] software, telco, retail, high-tech, [0:42] healthcare, and pharma. And the reason [0:43] why I put up this slide wasn't really to [0:45] give, you know, talk about mojo, but [0:47] ricotta was really about the importance [0:50] of processes in these companies. These [0:53] are not small companies that invent how [0:54] to do things. These are companies that [0:56] are incredibly rigorous at driving [0:59] operational excellence. I worked at [1:00] Amazon for three and a half years. Let [1:02] me tell you, they are rigorous at [1:03] driving operational excellence. And so [1:05] customers like this have really pulled [1:07] WCAT to get really good at meeting their [1:10] needs and delivering things that are [1:12] reliable and deliver process and [1:14] operational excellence. And I'm going to [1:16] come back to that theme. [1:19] So, Ricarda was founded with one core [1:21] idea and that is you have all of this [1:24] stuff needed to make your business run. [1:26] You have a CRM system and an HCM system [1:29] and you've got some business data and [1:30] some operational database and your legal [1:33] department does red lines with box and [1:35] and docuine right like the life is [1:38] complicated and work was built to be an [1:41] enterprise orchestration platform that [1:43] just connect all the things it could use [1:46] it business could use it developers [1:48] could use it has connections to more [1:50] than 12,000 applications you could run [1:54] all of your different processes through [1:55] it and help your business run more [1:58] efficiently and in a well orchestrated [2:00] way. And we call that orchestration. And [2:03] that's not the same as like dragging and [2:05] dropping a couple processes in in steps [2:07] in a process. It's really deeply having [2:09] great processes that run your business. [2:13] Processes are actually kind of what [2:15] differentiate your business. It's how [2:17] well you compete and how well you serve [2:20] customers and can you innovate and bring [2:22] new products to market fast. Do you give [2:24] your employees a great experience? Those [2:27] are processes. Those are not apps. They [2:29] cut across apps. They have human [2:32] components to them also. And in many [2:34] ways, they are the heart of how you [2:36] compete in your industry or serve your [2:38] customers or serve your stakeholders. Uh [2:42] processes also frequently almost always [2:44] cut across many many different systems. [2:47] And so even if you can see the state or [2:49] the data at rest in anyone's system or [2:51] the analytics to look back and see what [2:53] happened that doesn't actually tell you [2:56] the current state and health of that [2:58] process. [3:00] Okay. So I wanted to start with a big [3:02] problem that I see and I know there's a [3:04] lot of hype out there but agents all [3:06] this agent stuff generally broadly [3:09] speaking is still stuck at the edge. [3:12] Okay? It's in your productivity apps. [3:15] It's helping you autocomplete emails. [3:17] It's um uh uh guessing at what your [3:20] customer service intent might be to try [3:23] to route you to the right person. And [3:24] it's really really wrong a lot and often [3:27] scoped wrong. And I think all of us have [3:30] these big ideas of what amazing agents [3:32] could be that are deeply deeply capable. [3:35] Instead, they're focused on small tasks. [3:38] They tell you they're right and they're [3:39] sure when when they're really wrong and [3:42] the confidence is completely misplaced. [3:45] A lot of chatbot stuff where like you're [3:48] tippy the user's tippytapping and [3:50] driving at every step. It doesn't really [3:51] seem very autonomous if you're pushing [3:53] along like a toddler up the stairs, [3:55] right? So, uh, security, you know, there [3:58] are amazing experiments happening out [4:01] there, but if you really want an agent [4:03] to work with your corporate data, are [4:04] you really set up with a security [4:06] infrastructure to make sure the agent [4:08] can only see the right data on behalf of [4:09] the right users? Like, don't put payroll [4:11] in there, right? Because you don't know [4:12] what's going to happen. So, I think just [4:15] broadly as an industry, we're still kind [4:17] of stuck at the edge. Do do you see [4:18] this? I'm trying to be provocative here, [4:21] okay? But I think that's where we're at [4:23] and we got to go past that. And the [4:26] investment is incredible, right? Like [4:28] it's hard to keep up. This is just a [4:29] couple of months. GPT5, [4:31] uh, Claude 41, Opus from Enthropic, some [4:34] inc incredible new aogentic stuff from, [4:36] uh, from AWS, MCP, the A2A standard [4:41] looks interesting. About a thousand AI [4:43] conferences. I did a look. I'm in San [4:45] Francisco and there's 60 a week in San [4:48] Francisco right now. That seems like a [4:49] lot. And it's about a quarter trillion [4:52] dollars of capex that's been spent. So [4:55] all this innovation, all this money, yet [4:58] we're still kind of stuck at the edges. [5:02] Why? Like every every everything is [5:05] being applied. Why? Why are we stuck? [5:07] Well, it's really two things that jump [5:09] out to us. This is the first lesson I [5:11] wanted to share. That the agents really [5:13] lack enterprise context. We're [5:15] handcuffing them by not giving them the [5:17] business insights, the business data [5:19] they need to make good decisions. And [5:21] then the second thing is we don't really [5:24] trust them with agency. We have agents, [5:27] but we don't give them agency. We don't [5:29] let them do really important stuff, [5:31] right? We have to supervise. We don't [5:33] wire them to payroll directly. We have [5:36] to supervise them. And so we're not [5:38] giving them agency. So I that's all [5:41] negative, but let me let me give you a [5:43] picture of what the world can look like [5:44] and the kind of things we're doing with [5:45] our customers and learning a lot from. [5:47] Uh hi, how can we help you? Customer [5:49] service, my order was supposed to be [5:50] here Monday and I still don't have it. [5:52] And there's like levels and that basic [5:55] level and you've all got this. It's [5:56] like, hey, your order's on the way. It's [5:58] in LA. Now, when you get that, is that [6:00] helpful? [6:03] Maybe. Like I know you didn't lose it, [6:05] but is it moving or is it is it sitting [6:07] still? When's it coming? Okay, your [6:09] order's been stuck in the port of LA. [6:12] So, I've just decided the company's [6:13] decided we're going to get you a backup [6:15] pair and we're going to ship it a [6:16] different way. That feels really good. [6:18] That feels like like more more advanced, [6:20] more proactive customer service, right? [6:23] But what a company really wants is [6:25] having this stuff worked into the core [6:26] of the enterprise. And the company [6:28] realizes that that that signal triggered [6:31] an inspection. And the company realized [6:33] that all 67 orders are stuck in LA [6:35] because of a tariff thing. We have a [6:37] structural thing. It's not a one [6:38] customer service issue we have to fix. [6:40] We have a structural thing with one of [6:42] our processes. That's the kind of world [6:44] we need to chase after. [6:46] How do you do that? So, first thing we [6:49] talk about enterprise context is [6:50] 360deree data. That's like just getting [6:52] your data at rest from your different [6:54] apps and databases like where's the [6:55] order? Did it get paid? Uh what's the [6:59] shipment status and the tracking number? [7:00] It's kind of data at rest, right? And [7:03] even better than that is understanding [7:05] to to to enhance that second level is [7:06] understand those endtoend processes. [7:09] It's in the process of the order [7:10] creation, the pick from the warehouse, [7:13] the tracking number created, the handoff [7:15] to the carrier. So you can see exactly [7:17] in the process where it's stuck, the [7:19] process which may be different than just [7:21] all the individual source systems and [7:22] this reason for this event that's [7:24] happening in the process. [7:27] And then what you want to get to is this [7:28] world of business signals. Now these are [7:30] really different. You see this 82% [7:32] on-time delivery, 70% revenue risk, 78 C [7:37] set. What are those? That's not data in [7:40] CRM, right? That's a computed live value [7:44] about the health of your business and [7:45] the health of your processes. You [7:47] probably don't even have a system where [7:49] that live lives live. You probably do [7:51] that as a many companies that is after [7:52] the fact and computed on an analytics [7:54] warehouse. But this is the business [7:56] signal that could make agents really [7:58] good. This is like, you know, if you [8:00] have a great employee in that warehouse [8:02] in LA and they're like, "Everything is [8:03] stuck. We got to find a workaround for [8:06] the company." That's like magic and we [8:08] need to build agents as an industry that [8:10] could do that kind of magic. [8:14] So, this other thing that comes up is AI [8:17] is stochastic. I had to like remember [8:19] how to say this word, had to look it up [8:20] what it meant. But this just means [8:22] unpredictable, right? like it sort of [8:23] statistically works sometimes and [8:26] worksish other times and totally fails [8:28] sometimes. But I think all of us that [8:30] have worked in enterprise IT and just [8:32] want to deliver things for customers [8:33] like we need predictability like work [8:36] working sometimes is a really bad fit [8:38] for most enterprise IT. That's not how [8:40] we train ourselves as CIOS and CTO's and [8:42] IT leaders. And that's actually really [8:46] interesting if you dig into it. What is [8:48] it that makes us humans really good at [8:50] doing some things predictably? So, we [8:52] have this example. I didn't bring my [8:53] water up. You pick up a glass of water. [8:55] Uh, we do it without thinking, right? We [8:56] pick a glass of water, drink the water. [8:58] But we didn't know how to do that as [8:59] toddlers, right? It's like wobbly and [9:02] we'd spill. Watch out for the danger [9:03] zone with the cups to the table. We had [9:05] to learn that. But once we learned how [9:06] to do it, it's predictable. It's [9:08] repeatable. So, this thing on the right [9:10] is a recipe inside. [9:13] We call them skills now because we've [9:15] made them available to AI and your [9:17] company can express how you want to [9:20] create invoices and you do it once and [9:22] it becomes the way your company creates [9:24] invoices. It's codified as a skill that [9:27] now your employees and your processes [9:29] can use and your agents can use. This is [9:31] a really powerful building block that [9:32] we've had really good success with. [9:35] And in fact, there are thousands of [9:37] skills across your organization and you [9:39] need them to be really proven like [9:41] provisioning access for a new employee, [9:44] uh creating an invoice, tracking a [9:46] payment, verifying credit, like this [9:48] stuff is the stuff that has to work. [9:49] There are thousands of these things have [9:51] to work. And if the AI is just guessing [9:53] at these things, they will be stocastic [9:55] and they often won't work. You want them [9:57] secure, transactional, resilient, and [9:58] explainable. That thing is important. So [10:01] this is how we're thinking about [10:02] building core agents with our customers. [10:04] the idea that core agents start with a [10:06] KPI in the middle. I saw this in an [10:08] earlier presentation too that they were [10:09] thinking about this and that's really [10:11] exciting. Start with a business KPI. [10:13] What is the thing that this agent is on [10:16] a mission to make better for your [10:18] company? Just like you'd have a great [10:19] employee, you know, they like what is [10:21] their mission? Who do they serve? What [10:22] do they what is really do they do? They [10:24] might have a job description aligned [10:25] with that. And then give them on the [10:28] left all the enterprise context they [10:29] need and on the right all those [10:31] enterprise skills they can actually take [10:32] action. And then you get this really [10:34] powerful combination of the sort of [10:36] thinking brain of the LLM with a clear [10:39] definition of what it's supposed to do, [10:41] access to the context it needs and then [10:44] ability to take actions that actually [10:45] make changes in the business and drive [10:47] things inside the business and you move [10:49] to this kind of core agent world KPI [10:52] driven great context real time. It's a [10:55] very different kind of world of agents [10:57] and I think broadly as an industry we [10:59] have a lot less experience in how to [11:01] build those things on the right. If you [11:02] were just like follow the LM and [11:04] framework stuff, there's a lot of stuff [11:05] in the right that you know you don't get [11:08] in the Udemy course, you know, like it's [11:10] it's a combination of the new LM things [11:12] with these kind of core enterprise [11:13] skills and putting them together in new [11:14] ways. So, uh, we released this product. [11:18] I'm going to pitch you on it. We [11:19] released product called work one. So [11:21] after 10 years, we completely rebuilt [11:23] our platform to be designed for this [11:25] intersection of stochastic and [11:28] predictable for that enterprise [11:30] integration to connect every system. We [11:33] connect to more than 12,000 [11:34] applications. We connect to every [11:36] database. We connect to onprem. If you [11:39] have EDI, if you have uh you need RPA, [11:42] we offer that in our platform. It's a [11:44] very complete platform. And then we [11:46] built aic orchestration on top of that. [11:48] So you can start building agents and [11:51] building KPI driven core agents that [11:53] drive really meaningful impact for your [11:55] business [11:58] and something we're releasing very soon. [12:01] We released the first version of this a [12:03] month ago and I'm showing a preview of [12:05] what we're releasing in a week at our [12:06] conference in Las Vegas. So you this is [12:08] the world's first showing of this is [12:10] something called Marcato Go uh the [12:13] complete [12:15] uh um dashboard ccentric approach to [12:18] managing a fleet of agents in your [12:20] enterprise. So this is VJ my boss our [12:23] CEO. This is his dashboard. But what [12:26] you're looking at on the left are not [12:28] static computed KPIs. You're looking at [12:31] agents sharing their status updates. The [12:34] lead rocket agent is our outbound lead [12:36] genen agent. It works autonomously to [12:39] drive a target against gets a target set [12:41] of new leads. It looks at signal. It [12:45] identifies opportunities to connect with [12:47] a customer. It schedules emails out to [12:50] customers, prospects, and then it routes [12:53] the responses to the SDR team. It also [12:56] proactively designs new experiments for [12:58] better kinds of signal to act on or new [13:01] outbound emails to create. I I can give [13:03] you an example. Uh if someone is a [13:06] Ricardo customer and they like us and [13:08] they move to another company, that is a [13:10] really good person to reach out to the [13:12] future because they might want to again [13:13] bring to their new customer uh to the [13:16] new company they work for. And these are [13:17] the kinds of things that agents can [13:19] experiment with on your behalf [13:21] autonomously driving towards that KPI. [13:24] Super excited about that. So again, I [13:26] just want to communicate this is this is [13:28] a really exciting set of capabilities [13:30] we're launching on our customers and [13:32] that is autonomous agents driving [13:35] against KPIs raising their hand when [13:37] they need help but it you know reasoning [13:41] cognition taking advantage of all the [13:42] context trying new experiments to push [13:45] the business forward. [13:48] Okay, MCP is a hot topic and I also [13:52] wanted to highlight we're like [13:53] Switzerland. we connect everything. So [13:55] many of you are looking at ways to use [13:57] this MCP standard. Everybody's heard of [13:59] MCP, I think it's safe, model context [14:01] protocol. It's kind of a new a new way [14:04] um a sort of a layer on top of APIs to [14:07] make it easier for LMS to sort of find [14:08] them and learn how to use them and take [14:10] advantage of them. But there's lots that [14:12] it doesn't do, including many of the SAS [14:15] companies that say MCP, we love MCP, [14:18] don't offer MCP out. They are very happy [14:22] to take in MCP from someone else, but [14:24] they're very resistant to sharing [14:26] business data inside that database or [14:28] inside that application out. So, we do [14:31] that. Worcato has a serverless, fully [14:34] managed secure MCP server. You can take [14:36] any connection, any recipe and build a [14:40] crafted MCP endpoint for you to use [14:42] inside some agent project or use inside [14:45] cloud.ai or chatgbt. This is really [14:47] taking off fast for us. People like [14:50] that. um they don't get like a flat MCP [14:53] in front of an app like a sauna with [14:55] like all these API calls doesn't doesn't [14:56] really help the agent get stuff done. It [14:58] actually overwhelms it but instead build [15:00] an MCP that's kind of designed for a [15:01] process or a task like managing an order [15:04] and there's like a couple things in the [15:06] life cycle of the order and those APIs [15:08] and verbs get exposed to MCP. So you [15:10] compose an MCP layer that makes sense [15:12] for the agents that you're building. [15:13] That's how we built some of the things [15:14] internally for for um our agents. So [15:18] fully managed remote MCP server. [15:21] It was fun. I looked back over the [15:22] course of the year and every single one [15:23] of them had said orchestration. And it's [15:26] easy to think that orchestration is [15:27] like, okay, MCP that gets me APIs and [15:30] then I need a little doodlely box thing [15:32] and I got some workflows. We're good to [15:34] go, right? Yeah. Okay. The company's [15:37] been doing this for 10 years. There is a [15:39] lot more to it. Like how actually do you [15:42] handle security? That's basically not in [15:44] the MCP protocol. it's out of band. Um, [15:48] how do you handle retries in [15:49] transactions? That's probably not in [15:52] your workflow automation doodle thing. [15:54] So, you know, this is a world where [15:56] there's just a lot of depth and [15:57] complexity. We work really hard to make [15:59] it easy. I would just offer the advice [16:01] here as you go from a pilot to scale. [16:03] You will run into lots of these issues [16:05] and they're solve technical issues, but [16:07] you probably are not going to want to [16:08] solve them on your own and find a good [16:10] platform that will solve them for you. [16:11] I'll just leave that as general advice. [16:13] uh uh things we've learned move moving [16:15] to scale. [16:17] Okay, that's workado one. It's been [16:20] incredibly exciting to get this out. [16:22] Many of the our customers are are here, [16:24] our thousands of customers are here at [16:25] this event and so we're incredibly [16:27] excited to launch all these enterprise [16:28] agents uh next week at our event and I [16:32] put a link there if anyone is interested [16:34] coming back to Vegas two weeks in a row. [16:36] I know the food and the weather are [16:37] spectacular, but um it is awesome what's [16:40] coming. And if you hit me up on [16:41] LinkedIn, I probably have a discount [16:43] code, but uh I would love to see you at [16:45] our WOW event and I'd love to learn [16:47] about how you discover Ricardo and where [16:49] you take it. Thank you. [16:51] [Applause]