HIve: A Governed Agentic Workflow Platform for Healthcare With Exact Sciences Corp

Hive：携手 Exact Sciences Corp 打造的医疗受管智能体工作流平台 Glenn Fung 介绍了 Hive，这是一种四层智能体架构，旨在将标准化执行框架与定制化领域需求相结合，同时落实严格的医疗合规监管与状态持久性。该平台将自主 LLM 智能体嵌入确定性工作流图中以约束非确定性推理，并通过 IntelliDoc 以及由智能体群仅花费 0.34 美元构建的自动化采购工作流等用例进行了展示。Fung 还分享了基于编程智能体进行构建的实用工程经验，指出超快速的代码生成将主要开发瓶颈转移到了测试验证上，并要求团队遵循严密的扩展与整合周期。 Exact Sciences 拓展进军拥有 15 万名员工的 Abbott 企业，凸显了建立集中且可复用平台的必要性，以避免各业务部门之间进行冗余且孤立的 AI 开发。 (0:04) 在医疗健康领域部署智能体 AI 需要严格的可审计性、安全性和可追溯性，以便在不削弱开发速度的前提下保障患者安全。 (2:15) LLM 推理本质上是非确定性的——在相同的运行条件下往往会产生冲突的结果——这要求架构必须将非确定性推理约束在严格的确定性安全护栏之内。 (3:59) 受监管的智能体工作流需要持久的状态恢复、防篡改的血统溯源、默认安全的数据隔离以及不容妥协的人机协同（human-in-the-loop）检查点。 (3:59) 僵化的商业套件和完全自研的定制开发都无法单独满足企业需求；Hive 通过将标准化编排与可定制的领域逻辑相结合，填补了这一空白。 (7:07)

Glenn Fung, Vice President of AI Development and Strategy at Exact Sciences, argues that enterprise healthcare AI requires a governed hybrid architecture where non-deterministic agentic reasoning is strictly bounded inside deterministic clinical decision graphs. Bounding LLM Non-Determinism in Healthcare Environments (3:59) Deploying autonomous agents within regulated healthcare environments introduces risks that standard enterprise software does not face. Foundation models are fundamentally non-deterministic; Fung observes that running identical inputs and rules across 100 trials can yield one outcome 70% of the time and an alternate result 30% of the time, an unacceptable variance in diagnostic and clinical operations. Consequently, systems must enforce strict state durability, complete tamper-evident audit lineage, protected health information (PHI) isolation, and model portability across short two-month model refresh cycles. To prevent safety failures, human oversight is treated as a mandatory architectural boundary rather than an optional feature. Regulated agentic platforms must guarantee that interrupted patient-related processes can resume from exact historical checkpoints without data loss, while automated guardrails prevent prompt injection and data leakage before autonomous actions reach production systems. Hybrid Deterministic Graphs and Four-Layer Governance (14:24) Enterprise AI implementations typically polarize between rigid off-the-shelf harnesses, which are limited to generic conversational tasks, and bespoke home-brew projects, which generate high technical debt and siloed security. Exact Sciences built Hive to unite these approaches using four layers: an Agent Runtime Layer for reasoning and budget guardrails, an Orchestration Layer for multi-step graphs, a Control Plane Layer for observability, and an enterprise Catalog Layer for reusable agent capabilities, such as tracking at-home Cologuard collection kits via the UPS API. Hive handles clinical necessity through a hybrid deterministic model. Because medical guidelines follow strict conditional logic (evaluating condition A before routing to condition B), Hive preserves an immutable, deterministic graph structure. Unstructured, probabilistic LLM agents are restricted to individual graph nodes to handle messy inputs—such as legacy faxes and scanned PDFs in the IntelliDoc document pipeline—before returning structured data into the deterministic downstream path. Modular Execution Contracts and Model Cost Optimization (20:03) Hive isolates workflow authoring from underlying infrastructure through standardized execution contracts running on Synapse, Exact Sciences' AWS-hosted cloud platform. Workflows can execute interchangeably across orchestrators like Argo Workflows or native Kubernetes without altering application logic. Automated human-in-the-loop gates allow reviewers to inspect intermediate states and validate low-confidence extractions before payloads export to core operational databases. Cost governance is a core operational requirement when processing thousands of daily medical records. While Hive utilizes larger foundation models for initial exploration and workflow authoring, the engineering roadmap focuses on transitioning mature, high-volume production tasks to smaller, specialized fine-tuned models. As historical platform executions accumulate, task-specific models maintain extraction accuracy while substantially reducing ongoing inference expenses. Testing Bottlenecks in Rapid Agentic Code Generation (22:58) Hive Workflow Studio leverages a specialist agent swarm—comprising Architect, Builder, and Reviewer agents—to translate natural language descriptions into production-ready pipelines. In a live enterprise procurement demonstration, this multi-agent architecture generated a 9-step document workflow that reused seven catalog components, synthesized two new tools, verified PII safety, and executed a 6-page contract upsert at an authoring cost of $0.34. However, building Hive revealed an operational shift in software development: writing code has become cheap, turning rigorous test design and safety evaluation into the primary engineering bottlenecks. Fung notes that unconstrained code generation tempts teams into feature bloat without business justification. Sustainable development requires alternating between rapid expansion phases and disciplined consolidation phases to prune, standardize, and maintain architectural safety. While agentic swarms can compress cross-functional workflow development from months into days, Fung emphasizes that long-term viability depends on enforcing deterministic boundaries around LLM reasoning, validating business utility over raw code volume, and migrating high-volume pipelines to smaller, cost-effective models.

Exact Sciences人工智能开发与战略副总裁Glenn Fung指出，医疗行业部署AI智能体必须兼顾严格监管与敏捷创新。为此，团队研发了Hive受治理智能体工作流平台，通过将非确定性大模型限定在确定性图结构中，实现合规、可复用且具备人类监督的企业级自动化。 医疗AI的监管刚性与架构范式整合 (3:59) 在医疗健康与临床诊断场景中，智能体系统必须满足极高的合规标准，包括完整的状态持久化与故障恢复能力、不可篡改的端到端全生命周期审计日志、隐私健康信息隔离与注入攻击防护，以及不可妥协的人类在环（Human-in-the-Loop）监督机制。大语言模型内在的非确定性构成了显著临床风险。Fung指出，在100次相同输入和规则的测试中，模型可能70%的时间输出一种结果，另外30%输出另一种结果，这种偏差在医疗业务中无法接受。因此，系统必须将非确定性输出严格限制在确定性边界之内。 传统企业开发通常在两种范式间二选一：一种是开箱即用但场景受限的“预装框架（The Harness）”，主要处理通用问答；另一种是完全定制但导致孤立烟囱和高技术债的“自研系统（The Home Brew）”。Hive平台作为二者融合的统一架构，旨在提供标准化的基础设施与通用运行时，同时支撑复杂的专业领域逻辑，消除组织内部重复建设，确保每项开发均默认满足监管要求。 Hive四层分层架构与混合确定性执行 (9:58) Hive构建在运行于企业内部AWS Synapse云平台的“治理基座”之上，从下至上划分为四个解耦层：智能体运行时层（管理推理、动态护栏、上下文记忆与Token预算）、编排层（管理多智能体协作、Argo多步骤图与任务调度）、控制平面层（集中监控指标、执行日志、审批流与策略执行）以及资产目录/市场层。在目录层中，团队将诸如包裹追踪API等能力封装为标准化智能体供全公司复用，避免重复集成。 针对医疗必要性评估等临床逻辑，Hive采用了混合确定性架构。医院政策通常遵循严格的分支条件（如满足条件A则评估条件B，进而转入流程C），若交由非受控的LLM处理会导致极高风险。Hive的做法是由系统维护确定性的工作流有向无环图（DAG），仅在特定图节点嵌入自主智能体，专门负责从非结构化病历中提取诊断指标或临床观察。提取出的结构化数据随后返回确定性图执行后续分支路由，实现了基础设施无关的稳定执行契约。 智能文档处理与多角色协同降本 (16:46) 在实际落地中，Hive支持了名为IntelliDoc的非结构化医疗文档处理系统。该系统针对纸质扫描件、PDF以及传真等非标准化格式，执行摄取、智能体提取与分类、人机协同验证，以及输出合规JSON至下游电子病历系统的全流程。低置信度或高风险结果会被自动分流至人工复核队列。业务人员无需编写代码，只需声明数据输入、目录工具、输出模式及人工审批条件，平台便会自动注入弹性伸缩、状态恢复与权限审计能力。 面对每日数千份医疗文档的吞吐量，推理成本控制成为核心指标。Hive在初期实验阶段调用大型前沿基础模型，待业务流程成熟并积累历史运行数据后，系统会过渡至轻量级的任务微调小模型。由于编排层与底层模型解耦，这种模型替换无需重构上层业务逻辑，在保持准确率的同时大幅降低了推理支出。 编码智能体引发的研发节奏与质量重构 (22:58) 在利用编码智能体于数月内快速搭建Hive平台及业务用例的过程中，工程团队总结了软件开发范式的根本变化。由于AI生成代码的边际成本骤降，代码编写已不再是瓶颈，严格的测试设计与安全性验证反而成为了制约进度的主要环节。同时，极低的开发门槛容易诱发团队在缺乏明确业务需求的情况下盲目扩张无用功能，导致代码量激增而产品实际价值未增。 Fung提出，可持续的智能体辅助研发必须在“扩张”与“收缩”之间周期性交替。团队在快速原型设计与功能激增阶段后，必须主动暂停，进行代码清理、架构精简、标准化沉淀和技术理解，待基准稳固后再开启下一轮功能扩张，以此防范数千行失控代码带来的技术债务。 Hive工作流工作室的多智能体采购实战 (26:42) 在采购自动化用例演示中，Hive展示了通过专职智能体集群协同构建工作流的能力。面对持续流入的企业单据，业务需求要求仅提取供应商协议、核实采购支出并更新合同库以防止重复记录。系统没有采用单一庞大模型，而是调度了三个细分角色：架构师智能体将自然语言提示词转化为包含9个步骤、3个条件分支（供应商路径、归档路径与人工审核路径）和1个审查关卡的DAG；构建者智能体扫描目录后复用了7个现有组件，并自动合成了缺失的供应商字段提取工作器及防重复写入工具；审查员智能体则验证了组件的隐私安全与合规护栏。 整个工作流开发耗资仅0.34美元，几分钟内完成了以往需要数周的工程设计。在随后的自动化实测中，一份6页的供应商协议被正确归类（置信度93%）并执行分支抽取，验证了日期与物料后成功更新合同数据库（更新1行，0重复追加）。 Exact Sciences展示了Hive平台如何通过解耦分层、确定性图约束与多智能体装配，化解医疗场景下大模型非确定性与合规安全之间的冲突。不过，演讲内容基于内部平台Synapse及特定业务演练，如何在全组织数千名员工甚至并购后15万人的庞大规模下长期维持测试质量与代码治理，仍是有待持续验证的运营挑战。

## Background 本视频由 Ai4 频道于 2026 年 8 月 19 日发布在 YouTube 平台上。内容由 Exact Sciences 的 AI 开发与战略副总裁 Glenn Fung 主讲，介绍了 Exact Sciences 构建的企业级受监管智能体工作流平台 Hive，旨在将 AI 安全、可复用、可观测地直接嵌入到医疗健康的核心业务流中。 ## TL;DR 演讲围绕医疗健康高度监管环境下落地 Agentic 工作流的挑战与工程实践展开。Glenn Fung 详细剖析了 Exact Sciences 为兼顾严格合规（可审计、可恢复、控制非确定性、人在回路）与敏捷复用而设计的 Hive 平台架构，展示了结合确定性图谱与智能体推理的混合模型以及医疗文档智能处理（IntelliDoc）元用例。此外，演讲探讨了重度依赖 Coding Agent 开发背后的测试挑战与“扩张与收缩”代码治理机制，并现场演示了通过自然语言由专家 Agent 集群全自动组装端到端采购合规工作流的全过程。 ## Atmosphere & style 演讲整体氛围轻松幽默且务实坦诚，Glenn 以风趣的自嘲与幕后花絮拉近与听众的距离，同时在探讨医疗技术架构时展现出极高的专业度与严谨克制。现场节奏从宏观合规痛点层层推进到底层架构拆解与代码工程反思，演示环节则以惊艳的多 Agent 协同落地将情绪推向高潮，带给观众一种既前沿硬核又极具实战落地参考价值的沉浸式技术分享体验。 ## Key takeaways - [02:15](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=135s) 在医疗健康等强监管场景下落地 Agent，必须兼顾极高的合规控制（全流程审计与安全防护）与高度的业务组件复用，避免各团队各自为政重复造轮子。 - [03:59](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=239s) 医疗监管环境对智能体系统提出了严苛要求，包括持久性与故障恢复、全流程调用审计、将 LLM 非确定性收敛至可控边界、人在回路（HITL）监管以及底层模型可移植性。 - [07:07](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=427s) 现有落地范式存在两极分化：通用套件（The Harness）交付快但难以应对复杂核心医疗流程，业务自研（The Home Brew）能满足深度定制但复用性差且存在合规风险。 - [09:58](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=598s) Exact Sciences 打造的 Hive 平台融合了通用套件与自研方案的优势，分为智能体运行时层、编排层、控制面板层及目录市场层四大架构，建立在统一的治理基底下。 - [12:01](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=721s) Hive 采用三层系统设计（客户端接口、中台控制平面与执行契约、底层可移植引擎与基础设施），确保支持多角色交互并实现无厂商锁定（No lock-in）。 - [14:24](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=864s) 医疗业务逻辑高度依赖确定性规则路径，Hive 采用“确定性图谱 + 局部智能体推理”的混合模型（Hybrid Model），兼顾业务合规性与非结构化数据处理的灵活性。 - [16:46](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1006s) 针对医疗行业充斥纸质单据、传真及扫描 PDF 的痛点，Hive 孵化了 IntelliDoc 元用例，支持业务人员声明式定义抽取需求并自动注入安全护栏与人机复核机制。 - [21:29](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1289s) AI 辅助编码极大加速了系统交付，但带来了测试与审查追赶不上代码生成速度的挑战，团队内部必须在追求交付速度与坚守代码质量之间取得平衡。 - [24:04](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1444s) 代码量绝不等于产品价值，研发团队需建立“扩张与收缩（Expansion and Contraction）”的代码治理循环，在探索新特性后主动踩刹车进行精简整合，防止代码库失控膨胀。 - [26:42](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1602s) Hive Workflow Studio 演示展示了由架构师、构建器和审核员组成的多 Agent 协作集群，能够根据自然语言需求秒级组装复杂业务流，并支持按需即时生成专用组件与防重写逻辑。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=4s) 开场与背景：从 Exact Sciences 到雅培的 AI 探索 演讲在会议最后一程轻松幽默的氛围中展开。演讲者 Glenn Fung（曾任 Exact Sciences 的 AI 副总裁）打趣提到，他原本想说服团队成员上台为大家表演杂技来活跃气氛，但最终未能成行，只能由他一人独自完成分享。他谈到，大约在三四个月前他就受邀来分享相关工作，但由于 AI 领域变化太快，在这几个月里涌现出了更多更具价值的实践成果，因此演讲内容和主题也随之进行了迭代更新。 在个人与团队背景方面，Glenn 介绍了自己此前所在的 Exact Sciences——即肠癌无创早筛产品 Cologuard 的制造商。随着公司近期被雅培（Abbott）收购，团队规模经历了从 7,000 人到 150,000 人的跨越式转变。Glenn 回忆起与老朋友的对话，感叹自己兜兜转转又回到了超级大企业体系中。本次分享的核心，正是总结并分享他在过去八个月中与 IT 团队及合作方在构建 Agentic（智能体）工作流平台过程中积累的真实经验与教训。 --- ### [02:15](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=135s) 医疗场景下 AI 智能体的双重诉求：强合规与高复用 在医疗健康领域落地 AI 和 Agent 与其他常规业务场景截然不同。正如李飞飞（Fei-Fei Li）在大会上所比喻的，当下的 AI 革命在影响力上堪比工业革命或人类掌握火的使用，Agent 展现出了极其强大的潜力；但正因其强大，在医疗场景下必须保持极致的谨慎。任何进入临床或医疗业务流的 Agent 必须接受严格治理，具备全方位的安全保障、可审计性与行为可追溯性。 与此同时，企业又必须赋能业务创新，避免在基础设施层面陷入重复造轮子的低效困境。在大型企业内部，往往存在数十个团队各自为政开发 AI 项目的现象，导致大量重复建设与资源浪费。因此，平台化建设的核心目标是在确保安全合规的前提下，提供高复用性底座，帮助各团队快速构建与敏捷迭代。 ```mermaid flowchart LR accTitle: 医疗健康领域 AI 平台的核心双重驱动力 subgraph Governance["医疗合规约束 (Governance)"] G1["全程可追溯与可审计"] G2["确定性控制与安全防护"] end subgraph Innovation["企业敏捷创新 (Innovation)"] I1["组件高度复用"] I2["统一底座，避免重复建设"] end Governance --> P["Hive 平台设计原则"] Innovation --> P ``` --- ### [03:59](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=239s) 医疗监管环境对 Agentic 工作流的关键诉求 针对“受监管环境（Regulated Environment）”的具体技术落地标准，Glenn 详细拆解了团队内部头脑风暴并确立的几项核心特征： - **持久性与故障恢复（Durability & Recovery）**：系统必须能够随时恢复到工作流中的任意历史状态并重新启动。若患者决策流程中途意外中断，系统必须支持无缝续跑。 - **全生命周期可审计（Auditability & History Records）**：必须持久化记录所有 Agent 的每一次调用、输入输出和完整交互历史。 - **控制非确定性（Bounding Non-Determinism）**：LLM 本质上具有非确定性，但在医疗场景中必须严格收敛其输出边界。在过去两年的实验中，团队发现即便设定了极其严密细致的规则，同一个 Agent 运行 100 次，可能出现 70 次是一种答案、30 次是另一种答案的情况——这种波动在医疗场景中是绝对不被允许的，平台必须将非确定性收敛至确定的可控边界内。 - **人在回路监管（Human-in-the-Loop Oversight）**：确保关键决策节点有真实专业人员参与把关与复核。 - **默认安全与隔离（Safe by Default & Isolation）**：内置防御机制，主动防范提示词注入（Prompt Injection）、受保护健康信息（PHI）泄漏及敏感数据暴露。 - **可移植性（Portability）**：基础模型每两个月就会发生更替，整个系统必须实现解耦，使团队能够随时插拔和替换底层模型及工具链。 - **自助服务（Self-Service）**：满足企业内不同背景使用者的差异化需求，实现低门槛的自主配置与应用搭建。 --- ### [07:07](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=427s) 现有落地范式的局限：通用套件 vs. 垂直自研 在企业落地 Agent 时，业界通常面临两种互有妥协的路线： ```mermaid flowchart TD accTitle: 现有 Agent 落地范式对比 subgraph Harness["商业套件范式 (The Harness)"] H1["封装模型、记忆与执行环境"] H2["开箱即用、上线速度极快"] H3["局限：仅适用于就诊预约、客服等窄域场景"] end subgraph HomeBrew["业务自研范式 (The Home Brew)"] B1["完全根据自身业务深度定制"] B2["支持如事前授权等复杂核心场景"] B3["局限：复用性差、规范不一致、数据治理风险"] end Harness -.-> Goal["Hive：融合二者优势，补齐能力短板"] HomeBrew -.-> Goal ``` - **通用套件范式（The Harness）**：将模型与工具、记忆和执行环境进行一体化封装。该方式交付速度快、易于试验，但往往局限于标准化商业场景（例如患者答疑、服务检索、在线预约排期等），无法应对深度定制化的医疗核心流程。 - **业务自研范式（The Home Brew）**：各业务团队从零构建专属解决方案（如检测项目与医保结算高度相关的“事前授权/Prior Authorization”复杂自动化流程）。此类方案虽契合定制需求，但各团队烟囱式开发导致复用性极低、管控标准不一，且存在巨大的数据安全合规漏洞。 --- ### [08:43](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=523s) Hive 的定位：融合优势并填补空白 针对上述两极分化的痛点，Exact Sciences 设计了名为 **Hive** 的智能体工作流治理平台。Hive 的核心北极星目标是将通用套件的高效标准化与自研方案的高自由定制性有机结合，同时补全传统方案在审计追踪、成本治理和严格管控方面的空白，提供一套统一、安全、企业级的 Agentic 自动化底座。 --- ### [09:58](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=598s) Hive 四层架构与治理基底 Hive 整体架构建立在底层的**治理基底（Governance Substrate）**之上，向上划分为四大核心分层： ```mermaid flowchart TD accTitle: Hive 平台四层架构概览 subgraph Layers["Hive 平台分层"] L4["4. 目录与市场层 (Catalog Layer) - 智能体资产沉淀与跨团队复用"] L3["3. 控制面板层 (Control Panel Layer) - 可观测性、调用审计与运营治理"] L2["2. 编排层 (Orchestration Layer) - 多智能体协同、流转编排与工作流定义"] L1["1. 智能体运行时层 (Agent Runtime Layer) - LLM 推理执行、预算控制、记忆与 Guardrails"] end Substrate["底层基础设施：治理基底 (Governance Substrate)"] Layers --- Substrate ``` - **智能体运行时层（Agent Runtime Layer）**：负责执行 LLM 驱动的核心推理，并在执行时动态注入合规护栏（Guardrails）、Token 预算限制与会话记忆管理。 - **编排层（Orchestration Layer）**：协调复杂的多步骤工作流，管理 Agent 间的通信协议，提供工作流可视化定义与编排能力，确保各任务顺畅流转。 - **控制面板层（Control Panel Layer）**：作为端到端的可观测性中心，负责实时监控系统健康度、追溯历史执行轨迹、提供复核审查界面，并嵌入日常运营治理。 - **目录与资产市场层（Catalog Layer / Marketplace）**：将开发成熟的 Agent 和能力模块封装为标准化资产上架供全公司复用，杜绝重复开发。 Glenn 以实际业务为例说明了市场层的作用：在 Cologuard 肠癌筛查业务中，公司需要为患者邮寄样本采集套件。团队将对接 UPS 的物流追踪 API 封装为一个独立的物流状态追踪 Agent 并发布到市场中。无论是患者服务平台、内部客服中心还是其他业务线，均可直接调用该 Agent 获取样本寄送与在途状态，真正实现了企业级能力的即插即用与高效复用。 ### [12:01](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=721s) Hive 平台的三层架构与核心组件 Glenn Fung 介绍了 Exact Sciences 的 Hive 平台的整体架构。整个系统被清晰地划分为三个层次：最上层的客户端交互界面、中间的执行控制层，以及底层的后台支撑系统。不同背景的用户可以通过多元化的接口与系统交互——非技术业务人员可以使用直观的 UI 界面，而工程与算法技术人员则可以通过 API 或 CLI（命令行界面）直接调用所需能力，甚至通过事件驱动（Event-triggered）机制触发各类系统集成任务。 ```mermaid flowchart TD accTitle: Hive 平台的三层架构与核心模块 accDescr: 展示从客户端层到控制平面与执行契约，再到底层工作流引擎与基础设施的三层架构。 subgraph ClientLayer ["客户端层 (Client Layer)"] UI["Web UI 界面"] API_CLI["API / CLI 接口"] Events["事件触发集成"] end subgraph MiddleLayer ["中台治理与执行契约 (Control & Execution)"] ControlPlane["控制平面 (Control Plane)<br/>人机协同 / 审计 / 可观测性 / 策略合规"] ExecContract["执行契约 (Execution Contract)<br/>标准化稳定接口 / 动作解耦"] end subgraph CoreBackend ["底层引擎与平台支撑 (Runtime & Infrastructure)"] WorkflowEng["工作流引擎 (如 Argo Workflows)"] PortableCore["可移植核心 (Portable Core)<br/>运行时 / 日志 / Agent 记忆模块"] BackingPlatform["底层平台支撑 (Synapse on AWS / K8s)"] end UI --> ControlPlane API_CLI --> ControlPlane Events --> ControlPlane ControlPlane --> ExecContract ExecContract --> WorkflowEng WorkflowEng --> PortableCore PortableCore --> BackingPlatform ``` 在架构的中层，**控制平面（Control Plane）** 承担了保障人机协同（Human-in-the-loop）、全链路审计与可观测性、以及策略治理的核心职责；紧随其后的是**执行契约（Execution Contract）**，它作为一个稳定的抽象接口，定义并规范了智能体（Actors）所需执行的所有操作。底层则是以 **Argo** 为代表的工作流编排引擎、负责智能体运行时和记忆（Memory）的**可移植核心（Portable Core）**，以及运行在 AWS 之上提供基础支撑能力的 **Synapse** 底座。Glenn 强调，这种“无厂商锁定（No lock-in）”与基础设施无关的设计，让团队能够随时替换底层的具体实现组件。 - **多接口适配**：支持业务 UI、技术人员专用的 API/CLI 以及系统级事件触发。 - **可插拔引擎**：工作流编排层以 Argo 为例，但可根据需要自由替换，保持架构灵活性。 - **Synapse 平台底座**：在 AWS 上运行的核心支撑平台，为整个上层架构提供安全策略与算力支持。 --- ### [14:24](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=864s) 医疗场景下的确定性图谱与混合智能体模型 在总结 Hive 现阶段的技术成就与演进过程时，Glenn 着重阐述了医疗领域最为关键的**混合模型（Hybrid Model）** 理念。在医疗健康领域，业务流程往往受到严格的临床与法规准则约束，必须保持高度的确定性（Deterministic）。例如，在评估“医疗必要性（Medical Necessity）”时，医院和保险机构有着极其严格且固定的规则路径，整体业务逻辑必须被组织为一个“如果发生 A 则执行 B”的确定性图谱。 然而，在这个确定性图谱的各个节点上，系统必须依赖智能体（Agents）从非结构化文本中灵活提取关键信息并作出断言。Hive 的混合模型正是将这二者有机结合——既让智能体在局部充分发挥其信息抽取和推理能力，又将其严格限制在确定性的全局工作流框架内。同时，平台实现了“默认治理（Governed by default）”，确保任何团队基于 Hive 构建应用时，都无需重复担心合规与安全问题。 --- ### [16:46](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1006s) 医疗非结构化数据痛点与 IntelliDoc 元用例 谈及医疗行业的实际痛点，Glenn 幽默而无奈地提到，尽管现代科技高度发达，但全球医疗系统中仍然充斥着海量的纸质单据、传真件（Fax）以及格式各异的扫描 PDF。“对于不熟悉医疗行业的人来说可能很难置信，但即便是像德国这样高度发达的国家，传真在医疗流程中依然非常普遍。”大量关键的病历数据都是以这种非结构化形式传递的。 为了彻底解决这一痛点，Hive 孵化出了一个名为 **IntelliDoc** 的元用例（Meta use-case）。一个典型的智能文档处理工作流包含完整的端到端生命周期： ```mermaid flowchart LR accTitle: IntelliDoc 文档智能处理流水线 accDescr: 涵盖文档摄取、解析分类、提取断言、人机复核到最终数据转换的标准流程。 A["文档摄取<br/>(传真 / 扫描件 / PDF)"] --> B["解析与分类<br/>(OCR / 格式识别)"] B --> C["智能信息提取<br/>(智能体抽取与断言)"] C --> D{"置信度校验<br/>(HITL 人机复核)"} D -- "存疑/需确认" --> E["人工审核员审批"] D -- "高置信/通过" --> F["下游数据转换<br/>(生成标准 JSON / API)"] E --> F ``` 非技术人员只需以声明式的方式描述“需要从文档中获取什么”、“可调用的工具”、“期望的输出格式（如特定的 JSON 结构）”以及“是否需要人工确认”，Hive 平台便会自动注入所有的安全防护、日志审计与合规机制，开箱即用。 - **多角色协作生态**：业务人员用自然语言定义需求，平台工程师配置系统参数，治理团队嵌入合规准则，审核人员（Reviewer/Approver）在必要时介入人机复核。 - **成本与算力优化考量**：Exact Sciences 每天需要处理数以万计的医疗文档。目前系统虽然使用大型团队模型，但未来正计划逐步迁移至经过特定数据微调的轻量化小模型，以大幅削减日常推理的计算成本。 --- ### [21:29](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1289s) 自助式工作流生成与 AI 辅助编程的实践权衡 在工作流构建效率方面，过去跨部门业务人员与领域专家（SME）设计一套复杂的业务流程往往需要耗费数周甚至数个季度；而现在，借助 Hive 内置的智能体辅助生成能力，用户只需通过自然语言描述需求，智能体便能在几天内自动推断并生成可视化的工作流图谱，甚至主动提示“我可以为你执行此流程”，大幅加速了业务落地。平台同时自动接管了底层的高可用弹性伸缩（Auto-scaling）、持久化执行（Durable execution）、审计追踪与版本回滚等复杂运维能力。 在分享团队内部研发经验时，Glenn 直言，如果不是全方位重度依赖**编码智能体（Coding Agents）**，团队绝不可能在短短数月内同时交付 Hive 平台及多个复杂的业务用例。但这种极速交付也带来了工程文化上的新课题与深层权衡： - **速度与质量的博弈**：智能体让写代码的速度呈指数级提升，但“代码质量必须被极其审慎、有意识地监督（Intentionally supervised）”。 - **团队内部的思潮碰撞**：团队内部激辩不断——激进派希望借助 AI 跑得更快，而保守派则时刻警惕代码的安全性与隐患。 - **测试范式的转变**：当 AI 能够瞬间生成成千上万行代码时，自动化测试与严格的人工审查成为了确保系统安全稳健的最后防线。 ### [24:04](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1444s) AI 编程时代的测试挑战与“扩张与收缩”代码治理 在 AI 辅助编程极大降低编码门槛的当下，开发团队面临着全新的工程平衡难题。过去编写代码占据了绝大部分工时，而如今代码生成极其廉价迅速，导致“严谨的测试设计与执行速度往往跟不上代码生成的飞速步伐”。测试和验证所需的时间与精力，反而可能超过了实际生成代码本身。与此同时，由于做新功能变得轻而易举，团队极易陷入“既然我们能做，那是否应该做？”的功能蔓延陷阱。盲目堆砌实验性功能不仅容易分散注意力，还会导致代码库无序膨胀，因此团队必须高度克制并深思熟虑地决定究竟开发什么功能。 ```mermaid flowchart TD accTitle: 代码库的扩张与收缩循环 subgraph S1["快速探索阶段 (Expansion)"] A["AI 辅助高速编码"] --> B["探索新特性与业务实验"] end subgraph S2["治理整合阶段 (Contraction)"] B --> C["叫停盲目增长"] C --> D["审视业务真实价值 (Product Value)"] D --> E["合并精简代码与固化核心功能"] end subgraph S3["下一轮演进"] E --> F["受控可控的健康代码库"] F --> A end ``` 演讲者强调，“代码量绝不等于产品价值（Code volume is not equal to product value）”。面对生成几千行代码只需片刻的现状，技术管理者必须建立起“扩张与收缩（Expansion and Contraction）”的负责任治理机制： - **扩张期**：利用生成式 AI 快速构建、探索并验证哪些特性真正契合业务； - **收缩期**：在代码库膨胀至难以理解和维护前主动踩刹车，对代码进行深思熟虑的精简与整合（Consolidation），保留核心价值功能，剔除冗余； - 只有将代码规模维持在健康、可管理的合理水位后，团队才能放心地开启下一轮功能扩张。 --- ### [26:42](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1602s) Hive Workflow Studio 演示：多 Agent 协作与采购协议自动化 在 Hive Workflow Studio 的现场视频演示中，平台展示了如何通过纯自然语言描述将复杂的企业业务流程自动组装为生产级工作流。演示选用了采购团队（Procurement）的高频日常痛点：系统接收源源不断的企业文件，要求自动识别并仅处理“供应商协议（Supplier Agreements）”，从中精确抽取指定字段、完成合规校验，最终将数据无重复地写入“合同支出表（Spend Under Contract Table）”，同时附带完整的审计追踪（Audit Trail），并在遇到不确定情况时引入人工介入（Human-in-the-loop）。 ```mermaid flowchart LR accTitle: Hive Workflow Studio 多 Agent 协作生成与执行流程 subgraph Agents["Hive 专家集群 (Swarm)"] direction TB A1["架构师 Architect<br/>生成 9 步图谱与分支"] --> A2["构建器 Builder<br/>检索资产目录并即时生成"] A2 --> A3["审核员 Reviewer<br/>校验 PII 安全与 Guardrails"] end subgraph Artifacts["产出与执行"] direction TB W1["工作流定义 (YAML + Graph)<br/>7 个复用 / 2 个新建"] --> W2["真实执行 (6页协议样本)"] W2 --> W3["结构化结果输出<br/>更新 1 行 / 0 重复 / 成本 $0.34"] end Agents --> Artifacts ``` Hive 在后台通过由多个专业智能体组成的协作集群（Swarm of Specialists）实现端到端构建： 1. **架构师（Architect）**：解析自然语言需求并规划出包含 9 个步骤、3 个分支（主处理路径、归档路径、人工审核路径）和 1 个人工审核门禁的完整结构。 2. **构建器（Builder）**：基于组织资产目录（Catalog）按能力语义检索，迅速匹配并复用了 7 个成熟的已有组件（4 个 Worker 和 3 个 Tool）。 3. **按需即时生成（On-the-fly Generation）**：针对缺失的 2 个专用组件——“供应商协议专有字段提取 Worker”与“合同支出表写入 Tool”，平台在工作流中实时生成了规范（Spec），包括提取字段并解析站点代码与置信度的 `Supplier extract`，以及基于供应商与物料进行合并防重写的 `Spend contract upsert`。在用户透明预览并批准后，这两个新组件被正式注册进企业资产目录，供后续其他团队直接复用。 4. **审核员（Reviewer）**：在正式运行前对工作流的 Guardrails 进行全方位把关，确认新能力符合 PII 数据隐私安全后签名放行。 工作流生成后同时提供版本化的 YAML 代码与可视化的 DAG 图谱。在加载一份 6 页真实合同文档后，系统以 93% 的置信度将其识别为供应商协议，顺畅完成提取、物料校验、价格与站点代码解析，并在合同支出表中精准更新 1 行记录且无任何重复。整个生成与执行仅用时数秒，开发阶段的总模型调用成本仅为 34 美分（$0.34）。 --- ### [31:29](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1889s) 结语与花絮：AI 配音演示与经验分享 演示结束后，演讲者向现场观众透露了一个幕后花絮：为了规避现场实操可能出现的意外翻车风险，团队决定采用录制视频展示产品交互。然而由于内部团队成员都不愿意亲自出镜录音当解说员，团队索性使用 AI 语音生成了全程旁白。 演讲者笑着向大家澄清与致歉：“这听起来可能有点像商业推销视频，但请相信我们绝对不是在卖软件。我们只是在这里与同行坦诚分享 Exact Sciences 在落地 Agentic Platform 过程中的真实实践和经验。”随后，演讲者感谢全场参会者并在热烈掌声中进入问答交流环节。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=0s) 真正的企业级影响力来自于将 AI 直接嵌入业务工作流中，而不是孤立的模型调用。 — Glenn Fung > [03:59](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=239s) 同一个 Agent 运行 100 次，可能出现 70 次是一种答案、30 次是另一种答案的情况——这种波动在医疗场景中是绝对不被允许的，平台必须将非确定性收敛至确定的可控边界内。 — Glenn Fung > [00:00](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=0s) 这不是另一个 AI 框架——这是企业基础设施。团队可以使用 LangChain、PydanticAI、LlamaIndex 或 Strands，而 Hive 则补充了编排、治理和可观测性。无需重写代码，即插即用即可扩展。 — Glenn Fung > [16:46](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1006s) 对于不熟悉医疗行业的人来说可能很难置信，但即便是像德国这样高度发达的国家，传真在医疗流程中依然非常普遍。 — Glenn Fung > [21:29](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1289s) 代码质量必须被极其审慎、有意识地监督。 — Glenn Fung > [24:04](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1444s) 严谨的测试设计与执行速度往往跟不上代码生成的飞速步伐。 — Glenn Fung > [24:04](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1444s) 既然我们能做，那是否应该做？ — Glenn Fung > [24:04](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1444s) 代码量绝不等于产品价值。 — Glenn Fung > [31:29](https://www.youtube.com/watch?v=7oOjxkOEzKY&t=1889s) 这听起来可能有点像商业推销视频，但请相信我们绝对不是在卖软件。我们只是在这里与同行坦诚分享 Exact Sciences 在落地 Agentic Platform 过程中的真实实践和经验。 — Glenn Fung ## Mentioned Glenn Fung, Exact Sciences, Abbott (雅培), Cologuard, 李飞飞 (Fei-Fei Li), Hive, Synapse, Argo Workflows, AWS, IntelliDoc, Hive Workflow Studio, LangChain, PydanticAI, LlamaIndex, Strands, UPS ## Video description Glenn Fung - VP of AI Development and Strategy, Exact Sciences Corp As AI adoption accelerates across Exact Sciences, the real impact comes from embedding AI directly into business workflows—not from isolated model use. Our enterprise AI platform is being built to make AI integrations secure, repeatable, observable, and production-ready. It provides GPU-enabled infrastructure, production-grade MLOps, continuous delivery, observability for non-deterministic systems, human-in-the-loop guardrails, and cost governance as composable platform primitives. At the core is Hive, the foundation of our agentic ecosystem and the standard for building, deploying, and managing AI-powered automation. Hive orchestrates multi-step workflows, enforces compliance guardrails, tracks AI spend, and delivers end-to-end visibility from first model call to final output. This isn’t another AI framework—it’s enterprise infrastructure. Teams can use tools like LangChain, PydanticAI, LlamaIndex, or Strands, while Hive adds orchestration, governance, and observability. No rewrites required—just plug in and scale. [Source](https://www.youtube.com/watch?v=7oOjxkOEzKY)

[0:00] Glenn Fung: 真正的企业级影响力来自于将 AI 直接嵌入业务工作流中，而不是孤立的模型调用。 [3:59] Glenn Fung: 同一个 Agent 运行 100 次，可能出现 70 次是一种答案、30 次是另一种答案的情况——这种波动在医疗场景中是绝对不被允许的，平台必须将非确定性收敛至确定的可控边界内。 [0:00] Glenn Fung: 这不是另一个 AI 框架——这是企业基础设施。团队可以使用 LangChain、PydanticAI、LlamaIndex 或 Strands，而 Hive 则补充了编排、治理和可观测性。无需重写代码，即插即用即可扩展。 [16:46] Glenn Fung: 对于不熟悉医疗行业的人来说可能很难置信，但即便是像德国这样高度发达的国家，传真在医疗流程中依然非常普遍。 [21:29] Glenn Fung: 代码质量必须被极其审慎、有意识地监督。 [24:04] Glenn Fung: 严谨的测试设计与执行速度往往跟不上代码生成的飞速步伐。 [24:04] Glenn Fung: 既然我们能做，那是否应该做？ [24:04] Glenn Fung: 代码量绝不等于产品价值。 [31:29] Glenn Fung: 这听起来可能有点像商业推销视频，但请相信我们绝对不是在卖软件。我们只是在这里与同行坦诚分享 Exact Sciences 在落地 Agentic Platform 过程中的真实实践和经验。

Glenn Fung: As AI adoption accelerates across Exact Sciences, the real impact comes from embedding AI directly into business workflows—not from isolated model use. Glenn Fung: This isn’t another AI framework—it’s enterprise infrastructure. Teams can use tools like LangChain, PydanticAI, LlamaIndex, or Strands, while Hive adds orchestration, governance, and observability. No rewrites required—just plug in and scale. [22:58] Glenn Fung: We write a lot of code, but is that code safe? [26:42] Glenn Fung: Nine steps, seven reused, two built new. Weeks of engineering in a single conversation. [31:29] Glenn Fung: I, you know, can be tricky, not optimistic. So, it kind of looks like a sales video, but it's not. We're not trying to sell anything. We're just sharing with you our experience. Glenn Fung: At the core is Hive, the foundation of our agentic ecosystem and the standard for building, deploying, and managing AI-powered automation. [22:58] Glenn Fung: Because We Can, Should We? Glenn Fung: Hive orchestrates multi-step workflows, enforces compliance guardrails, tracks AI spend, and delivers end-to-end visibility from first model call to final output.

[0:04] Well, welcome everyone and uh we are at the end  of the day and I was trying to convince my team [0:09] uh to do some acrobatics for you like this  morning to see we can the energy but that [0:13] didn't work out. So, with no acrobatics for me, so  it's just me. Um and so, one of the things that um [0:23] I was um approached like, let's say three or four  months ago about um sharing some of the work that [0:31] we're doing um at Abbott. Now, Abbott and I'll  I'll talk a little bit about that. Um and then [0:39] we thought about we were in the early stages  um well, we have been some experimentation with [0:44] this agentic workflow uh platform that we have and  uh we you might notice that there is a different [0:52] title because like everything on AI, we thought  and we there is something more interesting that [0:57] I want to show right now right in the last three,  four months. So, that's the reason for that. Um [1:03] Uh about me, um yeah, I I'm the VP of AI  at formally Exact Sciences uh which is um [1:13] a company is the maker Cologuard. I don't know  if that kind of rings a bell. Of course, for the [1:18] people who work there, it rings a bell. But for  anybody else, you've probably seen the ads um [1:23] and we have been recently acquired by Abbott,  which is a well, we went from being part of [1:28] a 7,000 people company to 150,000 company now.  So, we're big and I was telling an old friend uh [1:36] here that I was like, "Oh, we're very small for  a while which uh is an interesting, you know, I [1:42] I came from an old big company and now but now  we're big again." So, uh we're into that now. Um [1:47] So, the motivation again of the presentation is to  share some learnings that we have had um um in the [1:55] last, let's say, I'll say 8 months of or so with  some of uh uh my coworkers, part of the IT team [2:03] um at our Castle Diagnostics. So,  let me [snorts] let me get into it. [2:10] Uh I need to point here. Wait, where  do I need to point? There you go. [2:15] Okay. So, there's two motivations for for  what you're about to see, for this this [2:20] conversation we want to have. So, first is  um when we do AI in health care is different [2:28] than some other fields uh other places where  you do AI. And especially talking about agents [2:34] and how how a powerful tool it is that you saw  this morning uh Fei-Fei Li talking about like [2:41] you can argue that's comparable to the industrial  revolution, the invention of fire. Right now, [2:47] we have this very powerful tool that is AI. And we  have these agents, right? But for health care, we [2:53] have to be super careful, right? They have to be  Whatever we do with AI agents have to be governed [2:59] because you have to be be able to audit what  they do, have to be secure. It has to be We have [3:04] We have to be able to trace. And this going to  be a common theme around this presentation, so [3:09] I'm going to come back to that in several examples  and talk you about how high it is and all that. Um [3:15] but at the same time on on this on the other side,  we we need to enable innovation and we need um you [3:21] know, a way to execute these things very well and  we enable to move faster. So, we need a platform [3:28] or we need a way that have reusability. We don't  want to replicate rebuilding things again and [3:34] again. I've been in industries again big companies  where uh there are like 60 projects going on on AI [3:41] by different teams and uh there is a lot of  replicated um infrastructure and that's uh [3:49] leads to inefficiencies, as you can imagine. So,  I'm going to talk a little bit more about that um [3:54] in a little bit. Yeah. So, going to  the first part of what I was saying, [3:59] what it means to be regulated. And this is us  brainstorming and thinking for our environment, [4:06] what it means to be, again, regulated. What  do we want these characteristics to be? And [4:13] so, the first one is about we should be able to go  back to any states that we had in the getting to [4:19] adopted in our workflow in the agentic workflow  and being able to restart it. So, that's what [4:24] durability and recovery, right? Imagine that  we are in a workflow that involves a patient [4:29] and we're in the middle of some decision-making  process. We want to be able if something gets [4:34] interrupted, we want to be able to come  back to that, right? So, that's one. Um [4:40] Related to that is we should be able to audit  and always have records um of the whole history [4:48] of everything that every agent is doing.  And that's very important for healthcare. [4:52] The third one I'm very passionate about and I  think I'm going to talk about it in another slide, [4:57] but it's what if you're going to be  non-deterministic, how all these algorithms are, [5:02] LLMs, you have to be able to get it close uh and  bounded and in some way being be able to control [5:10] the decisions that the agent make. And this is  probably something that you have seen in different [5:15] talks around different days because that's  something that we're figuring it out, right? I [5:20] In my experience, we have done in the last 2  years where we're very excited about an agentic [5:25] framework, and then we run an experiment with the  same data and the same output for um even with a [5:35] a good set of rules very well defined, and then  we run 100 times and we get 70% time one answer, [5:43] 30% another answer. So, that's not  admissible in healthcare. We want to get uh [5:48] very non-deterministic so the deterministic answer  so we want to bound whatever is non-deterministic. [5:54] I think I'll talk about that a little bit  later. Two, human oversight, human in the loop, [5:58] very important always, right? And I was in a panel  right now talking about that and I say this is not [6:03] a new idea. So, we have been thinking about human  in the loop in uh and some of people work with me [6:10] in the past um it's it's kind of it's smirking  like I could see over there cuz I'm always [6:14] talking about how important it is this is and um  safe by default and isolation is, you know, we [6:20] we need guardrails for all the things bad things  that can happen, prone injection, PHI leakage, [6:27] um sensitive data, etc. And portability, um as  you can imagine everything's changing super fast. [6:35] You want to be able to pivot and and  plug in different parts of your system [6:40] at different times, right? You the trivial  one is the models. The models change every [6:45] 2 months. You want a new model. You should be  able to be enable to pivot your model, right? Um [6:52] and self-service is something very important.  So, different users are going to need different [6:56] things for your identity frameworks. And um being  able to do self-service is something that is [7:02] uh is very important there. I want to  frame this into there is two paradigms. [7:07] And none of them kind of really is ideal  for us in our health care enterprise. Um [7:16] although I want to say that some of the solutions  that we have falls into one or the other or mix [7:22] of both, right? Um in one side is the harness. So,  it's infrastructure that wraps a model with tools, [7:28] memory, and execution environment give you  everything that you need. Very useful, but [7:33] it could be narrow to a specific type of use case,  right? And that an example is a platform that um [7:41] host, let's say, uh, agents that communicate with  in a commercial platform that you host them there [7:48] and answer patient questions, look locate  services, and schedule appointments, and so on, [7:53] right? Uh, there are many commercial platforms  that do that that that could do that that in [7:57] a good way, but they only work for that kind of  use cases, commercial as I said, harness, right? [8:02] Um, the home brew is you have a a use case  that is very specialized for your domain and [8:06] your company and you have to build it. So,  you have to go and build this thing, right? [8:10] Um, everything everything builds their own and  um, so, an example for us and um, I was talking [8:19] a little bit about that in um, earlier today, um,  is about prior authorization. So, suppose that [8:25] you have a flow um, of of of different uh,  service that you provide to a patient and you [8:31] need to be paid by insurance, but you need prior  authorization, how you automate that. And that's [8:36] is very specific for the kind of services that you  provide or test that you provide to the customer. [8:43] Okay, and then the goal is the following. So, we  we we define Hive um, as the, you know, one of the [8:49] the the north of of building this is the aims to  unite use the clauses and the strengths of these [8:56] two approaches at the same time that tries to kind  of close the gaps that these two approaches have. [9:01] And add some things that they don't have.  And, um, I don't want to go into super detail [9:08] into any of those, but again, you see, for  example, the harness is standard execution, [9:13] is very fast to deploy for those use cases,  um, easy experimentation, but narrow scope, [9:20] um, and different things that you don't want um,  or that can be limiting, right? And the home brew, [9:25] again, is very good for your you can customize  it because it's something that you build from [9:29] a scratch, it has limited reusability, sometimes  different parts of organizations are to build it [9:35] different, or they incomplete controls, uh some  risk with data that you know not all of them [9:42] all the use cases are going to have the same kind  of uh rigor, and those are some of the problems. [9:47] So, and Hive again aims to close those gaps. Yeah.  [snorts] So, now kind of after the motivation, we [9:58] I want to say what is it, right? I'm  a little bit try to get more into [10:02] into the details of that. So, again, I think  we can summarize Hive four layers. And they [10:09] all live on the one foundational infrastructure  that we call the governance substrate, right? Um [10:17] the components are probably components that  you've been hearing again during these 2 days here [10:24] because it's a common theme, so it makes sense  to have them. So, one is a agent runtime layer [10:29] where it executes the LLM-driven reasoning. And  also kind of make sure that you have the right [10:36] guardrails, budget, memory, and all that  that you need to that. Orchestration layer [10:41] coordinates multi-step workflows, and you know,  making sure that the agents talk to each other, [10:48] kind of define the workflow that you can author  the workflow, that you can create workflows, and [10:54] and make sure that all that kind of  runs smoothly. The control panel layer [10:58] is kind of more your observability layer when  you monitor, recall, review, and you have the [11:04] operational governance embedded. Right? And  then you have this idea of a marketplace [11:10] that is [clears throat] the catalog layer. It's  like you build new capabilities and new agents, [11:14] you should be able to come to that marketplace  and reuse it, right? In the platform, [11:18] so you don't have to rebuild things. An  example of that for us is we have a lot of um [11:25] for example, we have patients that we  send a kit to take a sample for Cologuard, [11:30] and we have a UPS API that we can we  have agents that wrap that up and then [11:37] can give you any information at any time regarding  where your kid is shipped, where it is, is it [11:42] already going to get to you, is it already on the  way. So, that could be an example of an agent that [11:48] can live in the marketplace and they will can  be used in many different places, right? So, a [11:52] little bit of the view very high-level view about  the architecture, right? So, and and the way that [12:01] I wanted to show this is how it interacts, what is  visible to the users. What is kind of the middle [12:11] execution layer that it stops what is visible  but interact with the user and what is behind, [12:17] uh you know, in the back all the all the system.  And that's the way it's kind of split in three [12:23] different ways that you see there. So, the first  one is the clients so, the users of the system [12:28] can interact with the system in many  different ways. So, one being through a UI, [12:32] for example, and you're going to see an example  later on that happens, or through an API or CLI. [12:38] Like some members of my team, for example, can be  clients or customers of of the this architecture, [12:47] but they don't need a way to create workflows  because they are technical. So, they need [12:53] something else from the system, right? Through  the APIs and and something else or the CLI. Um [12:59] also can interact with all the system integrations  through events that trigger some of the things [13:04] that that we want to do. The control plane is  where we make sure that all the human oversight, [13:09] human-in-the-loop, um being able to track how  things happen, audit, um observability, and so on, [13:18] defined by the users and the human is there, uh  policies, etc. And then we have the execution [13:23] contract um that's an interface uh that everything  the actors uh perform what the actors want to do. [13:32] And you'll see a little bit more in the demo about  that. [snorts] Uh and then under that we have in [13:37] general the workflow engine that making sure that  the workflow one is created it could be executed [13:43] in a good way and in a reliable way. And I put  Argo here as an example of a workflow engine [13:49] that we use, but it could be any. And that's part  of the no lock thing, so you should be able to [13:54] use different things in different parts of the  of the architecture. The portable core, again, [13:59] is more about, you know, engine runtime, logs,  memory, etc. that agents need to operate. [14:06] And um the backing service is more like the  platform that hosts all that. Um and what [14:11] we put here policy nest because synapse is our  um is our platform that it right now lives in AWS [14:18] that has a lot of core functionality that supports  all these all through the top. To give a summary [14:24] a little bit of about what Hive is and what  we achieve so far. Again, in this journey um [14:32] we have defined a requirement framework that um  keeps evolving. There's a lot of conversations [14:38] about what we need and how we need it inside the  team. Um uh and define what what we think is the [14:46] most half and again this is going to be changing  um all the time um and this always evolving. Um [14:52] our kind of view of executing an agent with memory  and everything that we do in order to have uh good [14:58] analytic um solutions um powering the enterprise.  And uh this one is the one that I say I'm going [15:06] to come back to it. It's hybrid model, right? It's  how we how we make sure that we have uh ways to um [15:14] being deterministic at at the same time that we  let the agent sometimes do what they need to do. [15:19] Right? That um um I I'm very opinionated about  because especially for me for the medical domain [15:27] you want to have your workflows very well  defined, right? Like imagine a medical workflow um [15:35] I can give you an example. For example, medical  necessity is a set of rules that are very well [15:39] defined and they are they live in hospitals with  medical rules and that you can imagine a graph [15:46] that is deterministic. This happen if that  happen and this happen and this happen and is [15:50] very well defined. But sometimes, right? Um you  have agents that need to extract information to [15:57] do the assertions that you need in that graph.  Imagine the notes in the notes of that graph. [16:02] Right? So, you need to have to combine a  hybrid model when you leave agents do what they [16:06] what they need to do to extract the information  that you need, but at the same time they follow [16:11] this very structured workflow. Right? So,  that's that's what we mean about hybrid model. [16:16] Um again, governed by default. Um so, Hive is  going to support applications that are going to [16:20] be created by different people and we want them to  um uh to not to be worried about being compliant [16:29] and being safe with the application. And again,  what is a self-sustaining independent contract? [16:34] Again, it's uh um this is going to run behind a  stable interface and it's infrastructure agnostic. [16:40] You know, we build it in a way that we can  change different components at any time. [16:46] [snorts] Okay, now I'm going  to go into a use case. Um [16:51] so, one of the use cases here, as you can imagine  in [snorts] the healthcare domain uh there is a [16:57] lot of information and believe it or not in in  the in the healthcare system all over the world [17:02] uh there is a lot of information that comes  in the in paper. There is a lot of faxes, [17:07] there is PDF and that's still happening. So, uh  for people who is not familiar with healthcare, [17:12] yes, that happens a lot. So, systems talk a lot  about uh with on a structured data. Even some of [17:18] the nature medical records sometimes some of them  come in unstructured format. Germany for example [17:25] is very big some facts still for medical processes  and all that. And so you can imagine that this is [17:32] something that is very common, right? So you get  these stages when you need to ingest the document, [17:38] you need to process it and extract the information  and all this information but classify it. [17:45] You need to validate something sometimes. You are  not sure about the extraction and that's you know, [17:50] that's where human the loop comes in the in the in  the picture. And you need to transform the output [17:58] for whatever the output has to be for a down  downstream task. Like you need a JSON file with [18:04] these characteristics, whatever you need. So the  all those all those steps there that we're seeing [18:09] there are a very common in some of these workflows  that are information depending that comes from a [18:18] sign and a structure way. So Hive supports what  we call IntelliDoc which is a meta use case [18:24] for the for the for the genetic platform that we  have. Um Okay, let me move on here. [snorts] Then [18:33] again just to re-emphasize this right like for  IntelliDoc then we know that all these things [18:40] that we have in the platform getting here. So  inherits all these things. So what a user does is [18:46] depending on the user again. Let's say is a a user  that is not technical. He declares what it has [18:55] to be done with the information that we get. The  tools that are available. The output the format of [19:03] the output of the data and whether or not a human  must confirm this in the in the workflow. And then [19:10] Hive does it by default all these protections, all  these different things that I talked about before, [19:17] Hive comes and give you that for free. So,  you don't need to worry every time you build [19:21] a solution of this type. Hive  has it for you there. [snorts] [19:29] This This graphic is is representing in  the in the Intelligog /Hive workflow, [19:36] how different actors interact or different  users interact with the system, right? So, [19:41] one is the person who creates the automation,  describe, sometimes it's in plain language, [19:46] and you're going to see in the demo in a little  bit. The platform engineer puts some stuff, [19:51] some requirements that has to has to be inserted  into the platform. The governance team tells you [19:59] that you need to be safe and you need to be  aware of these things, and that was some of [20:03] the input was in the platform that way. And then  there's the execution layer that it reviews, human [20:09] in the loop that can happen, and that's where a  reviewer and approver comes in in the platform, [20:14] and then there is the output that we generate,  the outcome for a downstream task, right? [20:24] Again, this is kind of double clicking for this  one, and this is more like an example, a concrete [20:31] example, about how these different execution  contract connects with different different [20:39] versions of the orchestration engine, Argo,  in this in this example for Intelligog, or the [20:45] infrastructure, right now it's Kubernetes that we  have it in there, but it could be something else. [20:51] And different team models again, right now we use  different team models, but in the future could be [20:56] smaller models that are pre-trained as we learn  and capture data in the system, we're aiming to [21:02] to lower the cost and have a slower  models, because again, we process [21:07] thousands of documents a day. And it's very  important for us to have models that are are [21:13] that are not super expensive to run. So, this is [21:17] just telling you right now, for example, we're  using some models that are from team models, but [21:21] in the future we might use smaller models that are  more customized for the solution, right? [snorts] [21:29] And again, about self-service workflow  creation, I did double click into this. Um [21:35] You have to create our flow our flow from scratch.  That might take, depending on the user, that might [21:40] take a lot of time. The way that we have agents  supporting the work the workflow creation, [21:47] it can take, you know, days in order to figure  this out and automatically suggest what the [21:52] workflow should be by a description. Um and you  can, you know, compose very easily without knowing [22:00] much about, you know, how to handle different  formats for different workflow platforms. You [22:06] just tell it what you need and it will suggest  and put graphical representations of the workflows [22:11] and it would tell you, "Okay, I can execute  that for you." And make comments. And uh so, [22:16] this is the idea of that, right? [snorts] That  it would take days in order to figure that out [22:20] with with the appropriate SMEs and the appropriate  people from the business instead of, um you know, [22:26] quarters or weeks. Um Yeah, and then again,  every every self-author workflow um gets [22:36] supported by all these different things that we're  providing in the platform, right? Auto scaling, [22:41] uh durable execution, um audit trails, and  roll back all the access is taken care of [22:48] for for for the user and all the different  things that that you're going to need there. [22:58] Okay. So, one thing I wanted to share, um, here is [23:06] of course the only way that we could  do this to using coding agents, right? [23:10] Uh, there's no way that we could do this in a few  months. Um, not only the platform, but also the [23:15] use cases that we do if we don't rely on coding  agents, right? So, um, a little bit of learning [23:21] is just this is very useful. You can accelerate  delivery and again, everybody knows that by now. [23:28] Uh, but there is some other thing that we have to  be cautious about. Like, we know now that quality [23:33] is achievable, but you need to be intentionally  it has to be intentionally supervised. [23:38] You need to be very intentional about doing this  and we still having conversations in my team [23:43] whereas people who wants to move faster,  there's people who is more conservative. [23:47] What is the right way to do this, right? Uh, we  we write a lot of code, but is that code safe? [23:53] So, there are many different opinions and still  this is an evolving learning that we have even [23:57] internally in my team. [snorts] Right? Testing.  So, it's funny, right? Because now testing [24:04] is might take more time to figure out the writing  the actual code. Right? So, rigorous testing [24:11] might not be able to keep pace with the generation  of the code, right? So, that's something that [24:16] didn't happen before, right? Testing design and  and being being able to be rigorous with testing [24:22] is now it might take more time, more effort  than the actual code generation, right? [24:29] Um, something else from the first one that I  didn't say is like nowadays because writing [24:35] software is so cheap and so easy, um, we we can  explore and and create a lot of extra features. [24:44] Um, so the question is because we can, should we?  And how, you know, what features should we work [24:50] on? So, it can be a little bit distracting  to just just to highlight that. Uh, that's [24:54] something also that we need to it's good because  you experiment and you know what sticks, but also [25:00] uh it might distract you or or or kind of grow the  code more than you want. So, this is a little bit [25:05] of learning what to create when how to create and  now that we can do all these coding really fast. [25:11] Uh the other one is code volume is not a  even so product value. [snorts] And again, [25:15] this is is obvious, right? But as I was building  what the business needs is is what is important, [25:19] right? Cuz normally again, related  to the first one, normally because [25:23] you can you should build it. You should be very  intentional about that. [snorts] Um And again, [25:29] one thing that I've been thinking a lot lately  and there is recently there have been a lot of uh [25:34] um kind of conversation about this is like  being able to generate high volume code [25:41] is okay, but requires a lot investment in  well-thought consolidation. Um like I feel [25:48] that we in the last months we have been kind of  expanding and creating these code bases that are [25:53] really really big. But there is a moment to stop  because they might might may become unmanageable [26:00] and see we can consolidate. And then when we  consolidate and we keep the features that we need [26:05] and we make sure that we that the code is in  a reasonable size, then we can grow again. So, [26:10] for me it's more like expansion and contraction  and making sure that you're responsible and think [26:15] a lot about that. So, that's something that we  are in the middle of that journey, too. You know, [26:20] um and um and I think uh other people is thinking  the same way. Oh, I generate this code that has [26:26] thousands of lines of code and I keep going and  generating features and now I don't even know. [26:31] It's very hard to understand the code or what  things are there. So, it's time to stop, see the [26:36] functionality and maybe maybe consolidate a little  and then we can generate new things. [snorts] [26:42] Okay. So, now I'm going to show a video. Um the  video is a demo, a live demo of the user of this. [26:51] Um The Well, the video the video when it In Hive  Workflow Studio, the concept is straightforward. [27:00] You describe a process in plain language and Hive  assembles the whole workflow for [music] you. [27:07] It reuses the workers and tools  your teams have already built. [27:11] And when something doesn't exist yet, with  your approval, it builds that piece on the fly. [27:18] So, let's give it a real job. Something our  procurement team deals with all the time. [27:25] Procurement gets a steady [music] stream of  documents. And for this specific use case, [27:30] only the supplier agreements actually matter.  Each document has to be properly classified, [27:36] have the right fields pulled out, checked, and  written into our spend under contract table [27:42] with no duplicates and a full audit  trail. [music] Anything the system [27:47] isn't sure about goes to a human. Let's watch  Hive [music] turn that into a working system. [27:54] Hive doesn't do this with a single model. It runs  a small swarm of specialists, an architect to [28:00] design the flow, a builder to assemble it, and a  reviewer to check the work before anything ships. [28:08] The architect [music] goes first, turning our  description into a real structure. Nine steps, [28:13] three branches, [music] and one human in the  loop gate. And it works out all the branching [28:18] on its own. The archive path, the review  path, and the main [music] path through. [28:25] Then the builder takes over. And this is where the  catalog earns its keep. Instead of building from [28:31] scratch, it [music] searches what the organization  already has and matches by capability. [28:36] It finds seven [music] pieces it can reuse, four  workers and three tools already built and trusted. [28:44] But two things don't exist yet. A worker for  our specific supplier agreement [music] fields [28:50] and a tool to write to the spend under contract  table. Normally, a project stalls here for weeks. [28:56] Hive just generates both right in the flow. And  nothing new is ever created behind your back. [29:03] Hive shows you exactly what it wants to add.  Supplier extract, which pulls every field, [29:09] resolves our sites to their codes and scores  its own confidence. And spend contract upsert, [29:16] which merges on vendor and material, so it  updates existing rows instead of duplicating them. [29:22] You can read the spec, edit it, or approve. We'll  approve and both get registered to the catalog, [29:29] ready for the next team that needs them. Before  anything runs, the reviewer checks the guardrails, [29:36] confirms the new capabilities are  PII safe, [music] and signs off. [29:42] And just like that, the workflow is  built. On the left, the exact YAML, [29:48] versioned and yours to keep. On the right,  the graph assembles itself one step at a time. [29:54] Nine steps, seven reused, two built new. Weeks  of engineering [music] in a single conversation. [30:02] Now, let's run it on a real document. Hive  reads six pages, [music] classifies it as a [30:08] supplier agreement at 93% confidence,  and sends it down the supplier path. [30:13] So, the archive and review branches simply aren't  needed this time. It pulls the vendor, the parts, [30:19] the materials, the pricing, and the sites. Checks  [music] that the materials match and the dates are [30:25] valid, and merges the results into spend under  contract, updating one row and appending none. [30:32] The audit trail is written and we're done.  And here's what comes out the other end. [30:38] The output tab gives you the structured result.  Every extracted field, the pricing, the resolved [30:44] site codes, and the merge summary showing one  row updated [music] and nothing duplicated. [30:50] Clean, structured data ready for the next system  to pick up. Seven of nine steps, two branches [30:58] skipped. We can track the total cost of developing  this workflow, 34 cents, with a confidence of 93%. [31:07] So, that's Hive. You describe the process  and Hive assembles what already exists, [31:13] builds whatever's missing, keeps a human in the  loop, and runs the whole thing in real time. [31:19] And everything new it creates becomes reusable for  the next team. Hive, workflows assembled. [music] [31:29] Okay, yeah, a disclaimer here is like we  wanted to make a video for for a live demo, [31:34] and that was the best way not to risk it, you  know, like live demos are are pretty hard. But, [31:38] nobody wanted to be the narrator, all the  thing. So, we decided to do it with AI. And [31:45] I, you know, can be tricky, not optimistic. So,  it kind of looks like a sales video, but it's not. [31:51] We're not trying to sell anything. We're just  sharing with you our experience, so, you know, [31:55] um but, yeah, so that's it. I don't know, you have  any Let me see. Um that's me That's it for me. No, [32:02] no, no, no. Okay, there. Well, thank you, and  you know, we're going to be around. Do you [32:07] have any questions? Um and yeah, so thank  you for being here today. Yep. [applause]