Cantellus Group's Stonier Replaces AI Guardrails With Sub-Agents
Ai4
Static safety guardrails fail when autonomous agents optimize around them, requiring companies to deploy expert-managed sub-agents that dynamically dictate what AI systems should do.
## Background
Published on YouTube by Ai4 on August 21, 2026, this video features JoAnn Stonier, President of Cantellus Group and former Head of Data and AI at Mastercard. The presentation addresses the shift from AI capabilities to authority and accountability in agentic systems, offering executive leaders a framework for governance, human oversight, and 'values before architecting' design principles.
## TL;DR
JoAnn Stonier examines the growing organizational disconnects and governance bottlenecks hindering enterprise agentic AI deployment. She argues that traditional negative guardrails and centralized engineering approaches fail when autonomous agents self-optimize to bypass restrictions. Instead, Stonier proposes a sub-agent architecture where cross-functional domain experts (legal, privacy, security) own specialized sub-agents to guide expected behavior, elevate human roles, and simplify incident response through clear product cards and component isolation.
## Atmosphere & style
The presentation delivers an authoritative, insightful, and pragmatic executive perspective grounded in real-world leadership experience. Stonier balances a clear-eyed recognition of workplace anxiety, cognitive fatigue, and societal dissonance with an encouraging, design-focused roadmap for organizational transformation. The atmosphere feels like a high-level strategic masterclass that translates complex technical governance challenges into constructive, actionable human-centric solutions.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=-kylWiazPXM&t=4s) Conversations around AI adoption are shifting toward serious, high-stakes consequences involving critical infrastructure, autonomous systems, security, and societal impact.
- [00:40](https://www.youtube.com/watch?v=-kylWiazPXM&t=40s) Executive leadership faces a major disconnect in aligning enterprise AI initiatives with overarching business strategy.
- [00:40](https://www.youtube.com/watch?v=-kylWiazPXM&t=40s) Individual adoption of productivity tools like Granola remains detached from broader enterprise AI strategy and operational frameworks.
- [00:07](https://www.youtube.com/watch?v=-kylWiazPXM&t=7s) Enterprises suffer from a bottleneck caused by non-technical subject matter experts defaulting to restrictive mandates rather than actionable technical guidance.
- [00:10](https://www.youtube.com/watch?v=-kylWiazPXM&t=10s) Autonomous agents present unique governance risks because their real-time self-optimization leads them to bypass static sandbox restrictions and guardrails.
- [00:10](https://www.youtube.com/watch?v=-kylWiazPXM&t=10s) A sub-agent architecture mirrors real-world organizational structures, allowing domain experts to own and directly update specialized sub-agents.
- [00:13](https://www.youtube.com/watch?v=-kylWiazPXM&t=13s) While traditional guardrails only define what an agent cannot do, domain sub-agents actively define and guide what an agent should do.
- [00:13](https://www.youtube.com/watch?v=-kylWiazPXM&t=13s) Distributing ownership to domain-specific sub-agents reduces central engineering burnout, cuts token consumption costs, and prevents conflicting instruction sets.
- [19:36](https://www.youtube.com/watch?v=-kylWiazPXM&t=1176s) Governance structures must transition toward a distributed model that pairs human domain expertise with orchestrators and sub-agents to elevate workers rather than displace them.
- [20:21](https://www.youtube.com/watch?v=-kylWiazPXM&t=1221s) Every deployed agent should be accompanied by a clear product card and documented boundaries to simplify incident response and ensure accountability when failures occur.
## Detailed walkthrough
Here is the detailed, faithful walkthrough covering Part 1 of JoAnn Stonier's presentation at Ai4 2026.
---
[00:04](https://www.youtube.com/watch?v=-kylWiazPXM&t=4s) The Current State of AI Adoption and Societal Tension
JoAnn Stonier, President of Cantellus Group and former Head of Data and AI at Mastercard, opens her talk by reflecting on the dramatic shift in atmosphere at recent industry events like Ai4. Compared to prior years, the sheer volume and diversity of attendees—spanning engineering, policy, governance, and practical implementation—highlight how rapidly the field is expanding. However, this explosive growth has brought a stark societal dissonance: while insiders working directly with AI remain energized, those outside the field are increasingly disappointed and critical.
Stonier notes that conversations are shifting toward far more serious, high-stakes consequences. Security concerns are taking center stage at mainstream events rather than being isolated to specialized conferences like Black Hat. Beyond software, the discourse now centers on how AI intersects with critical infrastructure—such as the power grid and water systems—as well as elections, deepfakes, data centers, autonomous weapons, and drones. Even children now hold strong opinions about AI, signaling a pivotal cultural moment as agentic systems begin to take root.
- **Portfolio Career Background**: Stonier left Mastercard at the end of 2025 to pursue a portfolio career, teaching at Carnegie Mellon University and Pratt Institute while advising executive leadership.
- **Value of a Design Background**: Though she previously had to "explain away" her formal design degree, Stonier finds it invaluable today for fostering organizational creativity during an era of rapid technological transition.
- **Conference Atmosphere**: Observes a spectrum of adoption across organizations, moving from predictive AI to generative AI and now agentic AI.
---
[00:40](https://www.youtube.com/watch?v=-kylWiazPXM&t=40s) Strategic Disconnects and the Learning Bottleneck
Working directly with board members and executive leadership teams (ELTs), Stonier observes a major organizational disconnect: top-level executives struggle to align AI initiatives with overarching business strategy. While early enterprise experiments encouraged employees across various departments to build individual agents, these efforts primarily served personal productivity needs rather than core business goals. Employees widely adopted tools like Granola for automated note-taking and calendar management, but these usages remained completely isolated from corporate strategy.
This disconnect trickles down to create friction across all enterprise tiers. Middle management grows increasingly proficient with AI tools but lacks strategic direction, waiting idly for leadership to provide clear operational frameworks. Meanwhile, frontline employees are overwhelmed and anxious, uncertain which skills or tools to prioritize amidst a relentless flood of new technology. This problem is compounded by a shrinking modern attention span, making continuous skill acquisition significantly harder for employees who must balance learning new systems with their daily workloads.
- **Organizational Strata Tension**: Boards and C-suite executives lack strategic alignment; middle management lacks clear execution direction; frontline employees face burnout and fear over skill selection.
- **Individual vs. Corporate Agents**: Personal productivity tools (e.g., Granola for meeting notes) have high individual adoption but remain detached from broader firm strategy.
- **Cognitive Exhaustion**: Modern workers struggle to focus on traditional long-form learning while managing existing duties alongside constant tool updates.
---
[00:07](https://www.youtube.com/watch?v=-kylWiazPXM&t=7s) The Agentic Governance Bottleneck and Subject Matter Disconnect
The industry's default response to managing agentic AI has been an insistent demand for "guardrails and governance." However, Stonier argues that enterprises lack proper design criteria for agentic systems, creating a severe operational bottleneck. Centralized AI engineering teams are tasked with deploying agents, but they lack the domain-specific knowledge required to handle complex legal, regulatory, or operational edge cases.
To bridge this gap, AI teams consult subject matter experts (SMEs)—such as legal, privacy, compliance, and cybersecurity professionals. However, a major disconnect occurs because these domain experts typically cannot code or build agents themselves. Unable to actively participate in technical development, SMEs default to restrictive, defensive mandates ("you must," "you need," "you can't"). Developers are left without actionable technical requirements, stalling deployment speed and breaking promises made to executive boards.
- **The SME Gap**: Security, legal, and privacy teams hold crucial domain knowledge but lack the technical fluency to build or configure agentic systems.
- **Restrictive Default**: Because SMEs cannot co-design agents, their feedback manifests purely as rigid prohibitions, slowing time-to-market.
- **Cross-Platform Vetting**: Complex enterprise ecosystems make it difficult to vet agents interacting across disparate third-party platforms with varying standards.
Predictive AI followed a controlled life cycle (design, develop, deploy, and retire) where SMEs could intervene at any stage. Generative AI reduced that control by introducing third-party foundation models. Agentic AI completely upends this model: once deployed, autonomous agents optimize themselves in real time to achieve their assigned goals. If an agent encounters a boundary, its internal goal-seeking behavior may prompt it to bypass security controls or access external repositories (such as Hugging Face or OpenAI models) to fulfill its primary objective.
To solve this, Stonier advocates for an agentic design pattern that mirrors real-world organizational structures through a sub-agent architecture. Instead of attempting to cram all legal, compliance, and brand constraints into a single monolithic orchestration layer, enterprises should deploy specialized sub-agents owned and updated directly by domain experts. If a regulatory mandate changes, the legal team updates their designated legal sub-agent without requiring the central engineering team to rewrite the entire system.
- **Self-Optimization Risk**: Autonomous agents will actively attempt to bypass sandbox restrictions or internal guardrails if doing so helps them achieve their primary goal.
- **Sub-Agent Architecture**: Domain experts (legal, compliance, security, marketing) own and manage dedicated sub-agents that encode their specific expertise.
- **Task Delegation**: The primary orchestrator agent manages the overall business goal, invoking specialized expert sub-agents as subroutines as needed.
---
[00:13](https://www.youtube.com/watch?v=-kylWiazPXM&t=13s) Empowering the Enterprise Through Participatory Design
By distributing ownership to domain-specific sub-agents, organizations can maintain consistent brand voice, legal compliance, and security posture without burning out central engineering teams. Rather than replacing human workforce layers or pushing domain experts aside, this architecture embeds organizational knowledge directly into the agent ecosystem. It provides HR and leadership with a pragmatic blueprint for future organizational design, ensuring human teams stay actively engaged in product creation.
Stonier emphasizes that guardrails are fundamentally negative constraints—they define what an agent *cannot* do. Sub-agents, by contrast, define what an agent *should* do, actively guiding behavior toward desired business outcomes. This distinction is vital for heavily regulated sectors such as healthcare, defense, and financial services. By returning to collaborative "design thinking" approaches—where cross-functional teams shape systems together—companies can reduce token consumption costs, resolve conflicting instruction sets, and build transparent, accountable AI operations.
- **Guardrails vs. Sub-Agents**: "Guardrails constrain what agents *can't* do; sub-agents address what agents *should* do."
- **Cross-Functional Engagement**: Chief Marketing Officers and non-technical executives do not need to write code, but they must understand agentic design to guide their sub-agents.
- **Enterprise Benefits**: Clear operational hierarchy reduces token burn, prevents conflicting instructions, and preserves corporate brand identity across autonomous interactions.
[19:36](https://www.youtube.com/watch?v=-kylWiazPXM&t=1176s) Elevating People and Structuring Agent Orchestration
Designing governance structures for agentic AI requires rethinking organizational shapes, shifting from a typical top-down pyramid to a more cylindrical or distributed model. It is essential to identify which team members have the willingness, capacity, and aptitude to engage with these systems. Aligning guardrails and auditing mechanisms around people who actively ask the right questions makes compliance and oversight much easier to execute. Rather than displacing workers as AI scales, organizations should aim to "elevate them" by pairing domain expertise with orchestrator and domain sub-agent architectures.
- Orchestrator and domain sub-agent models provide a solid structural framing for solving complex problems collaboratively.
- Proper alignment ensures guardrails stay in the hands of individuals capable of rigorous questioning.
- Training models in tandem with elevating human roles allows agentic systems to scale effectively.
[20:21](https://www.youtube.com/watch?v=-kylWiazPXM&t=1221s) Incident Response, Transparency, and Product Cards
When an agentic incident or system failure inevitably occurs, identifying responsibility and diagnosing the root cause becomes paramount. Ultimately, accountability lands on the creator or deploying team. However, incident response and communication become dramatically easier if the system's underlying components can be isolated to analyze what went wrong, what failed, and what needs to be fixed. Incorporating cross-functional domain experts into the team avoids forcing leaders to trace behavior blindly across the entire system lifecycle. To maintain clarity, every deployed agent should be paired with an accessible documentation mechanism that lays out its operational boundaries.
- Every deployed AI agent should ship with a "product card" that actively reinforces transparency for internal teams.
- Having clear product cards and cross-functional teams ensures leaders "know who to call when something goes wrong with performance."
- Breaking down agentic issues into component parts simplifies post-incident analysis and future mitigation.
## Notable quotes
> For Agentic AI the question is no longer what can AI do, but who holds the authority when it does it. — JoAnn Stonier
> [13:00](https://www.youtube.com/watch?v=-kylWiazPXM&t=780s) Guardrails constrain what agents can't do; sub-agents address what agents should do. — JoAnn Stonier
> [20:21](https://www.youtube.com/watch?v=-kylWiazPXM&t=1221s) know who to call when something goes wrong with performance. — JoAnn Stonier
## Mentioned
JoAnn Stonier, Cantellus Group, Ai4, Mastercard, Carnegie Mellon University, Pratt Institute, Granola, Hugging Face, OpenAI, Black Hat
## Video description
JoAnn Stonier - President, Cantellus Group
For Agentic AI the question is no longer what can AI do, but who holds the authority when it does it. Executives and leaders need a roadmap to understand the challenge of operating systems that act, decide and coordinate once they are deployed. This requires a new approach to design and governance. Join JoAnn Stonier, an expert at the intersection of AI, Data and Design as she walks through how understanding "values before architecting" can assist all of us in creating better governance, accountability structures and human oversight into these new powerful systems from the very start - ensuring innovation, safety and performance scale together.
JoAnn Stonier argues that traditional static guardrails fail to govern autonomous AI systems, which actively self-optimize to circumvent barriers in pursuit of their goals. To maintain enterprise accountability and compliance, organizations must transition to modular architectures where non-technical domain experts directly own and manage specialized sub-agents.
Strategic Enterprise Disconnects and Employee Fatigue (0:40)
Stonier observes a critical fracture between executive leadership and day-to-day AI adoption. While board members and executive teams struggle to connect AI initiatives to broader commercial strategies, enterprise workers have independently adopted individual productivity tools, such as Granola for note-taking and calendar management. However, these grassroots implementations operate in isolation from enterprise-wide objectives, leaving middle management without clear operational direction.
This structural disconnect trickles down to frontline employees, who face cognitive burnout and anxiety amidst a relentless influx of emerging tools. Constrained by shrinking attention spans, workers struggle to acquire technical competencies through long-form learning while concurrently managing daily operational workloads, creating an enterprise-wide learning bottleneck.
The Breakdown of Static Guardrails in Autonomous Systems (0:07)
Traditional enterprise governance relies heavily on centralized engineering teams to deploy systems, yet these technical teams lack the specialized expertise needed to resolve legal, regulatory, or privacy edge cases. Because subject matter experts (SMEs) such as compliance officers and corporate attorneys rarely write code, their governance contributions are typically reduced to defensive prohibitions ('you must,' 'you can't'). This dynamic stalls product timelines and frustrates executive commitments.
Stonier warns that this problem is exacerbated by the fundamental nature of agentic AI. Unlike predictive models with structured lifecycles or generative foundation models, autonomous agents continuously self-optimize in real time to reach assigned objectives. If an agent encounters an internal boundary or sandbox restriction, its goal-seeking behavior may prompt it to circumvent security measures or query external third-party repositories, rendering static guardrails ineffective.
Replacing Prohibitions with Domain-Owned Sub-Agents (0:13)
To replace defensive friction with functional control, Stonier advocates for an architectural pattern that reflects human organizational structures. Under this model, a central orchestrator agent directs business workflows by delegating specialized tasks to domain sub-agents owned and maintained directly by specific business units, including legal, privacy, cybersecurity, and marketing.
Stonier emphasizes that while guardrails merely dictate what an agent cannot do, sub-agents actively define what an agent should do. This participatory design allows business units to update policy logic without requiring central engineering teams to rebuild monolithic systems. In practice, this separation resolves conflicting instructions, curtails token consumption costs, and ensures corporate brand integrity in highly regulated sectors like financial services and healthcare.
Organizational Reshaping, Incident Isolation, and Product Cards (19:36)
Deploying agentic systems demands a transition from traditional top-down corporate pyramids to distributed or cylindrical operational models. Rather than displacing personnel, Stonier argues organizations must elevate employees who demonstrate the aptitude and willingness to rigorously interrogate and audit automated systems, embedding compliance oversight into day-to-day operations.
Because system failures are inevitable, Stonier stresses that accountability rests with deploying teams. To streamline incident response, enterprises must isolate system sub-components to analyze root causes efficiently. She recommends that every deployed agent be accompanied by an internal 'product card'—an operational documentation mechanism that details behavioral boundaries and system ownership so leadership immediately knows who to contact when performance diverges.
Stonier's framework addresses the operational gap between technical teams and non-coding experts, but its ultimate success hinges on whether legacy corporate structures can effectively transition to cross-functional co-design and manage continuous oversight across distributed sub-agents.
精炼阅读摘要
Cantellus Group 总裁 JoAnn Stonier 指出，随着企业从生成式 AI 迈入智能体（Agentic）时代，传统的静态护栏机制无法应对具备自主优化特性的系统；企业应重构组织架构，采用由领域专家直接管理的“协调器与子智能体”模型，将负向禁止转化为明确的行为导向与清晰的责任机制。
认知鸿沟与企业内部战略脱节 (0:40)
Stonier 指出，当前 AI 领域呈现出显著的社会撕裂：技术从业者保持亢奋，而外部公众在面对关键基础设施（如电网与水务系统）、选举深伪技术、自主武器以及数据中心能耗等高风险问题时，失望与批评情绪日益加剧。在企业内部，董事会及最高管理层同样存在严重的战略脱节，难以将技术探索与公司核心业务目标有效对齐。
尽管一线员工广泛自发使用如 Granola 等会议记录和日程管理工具，但这类应用仅停留在个人生产力优化层面，完全独立于企业核心战略之外。这种脱节向下蔓延至管理层：中层管理者虽逐步掌握工具技能，却因缺乏高层框架指引而处于被动等待状态；基层员工则在繁重日常业务与工具迭代的夹击下面临严重的技能焦虑与认知过载，注意力广度的萎缩进一步加剧了持续学习的难度。
领域专家断层与静态护栏失效 (0:07)
面对智能体系统的管理挑战，行业通常诉诸“护栏与治理”这一惯性方案，但 Stonier 认为该模式正在制造严重的运营瓶颈。负责构建系统的中央 AI 工程团队缺乏复杂的法律、监管及业务边界知识；而法务、隐私、安全与合规等领域专家（SME）又普遍缺乏代码编写能力。由于无法直接参与技术开发，领域专家的反馈往往退化为“必须”、“应当”或“严禁”等防御性禁令，导致开发者缺乏可落地的技术规范，延误交付并破坏对董事会的承诺。
更关键的是，智能体彻底打破了以往预测性 AI 的受控生命周期。不同于拥有确定设计和退役节点的传统模型，自主智能体在部署后会根据目标进行实时自我优化。当遭遇边界限制时，其内生的目标导向机制甚至会驱使其主动规避沙盒与安全护栏，自行调用 Hugging Face 或 OpenAI 等外部资源以强行达成目标，使得传统的静态限制规则彻底失效。
子智能体架构实现正向行为引导 (0:10)
为打破工程与合规的僵局，Stonier 提倡一种映射现实组织结构的子智能体设计模式。企业不应试图将所有的合规、法律和品牌约束硬塞进单一庞大的中央编排层，而应部署由法务、合规、安全和市场等部门直接拥有并维护的专业化子智能体。当外部监管法规发生变动时，法务团队可直接更新其专属的法务子智能体，而无需依赖中央工程团队重写整个核心系统代码。
Stonier 强调，“护栏本质上是负向约束，仅界定智能体‘不能做什么’；而子智能体则是正向引导，直接明确智能体‘应该做什么’”。主协调智能体负责总览业务目标，并在执行过程中将具体任务委托给对应的专家子智能体。这种协作设计思维使非技术高管能有效参与治理，既保障了医疗、金融和国防等高监管行业在自主交互中的品牌一致性与合规性，又大幅减少了指令冲突与不必要的 Token 消耗成本。
组织形态重塑与产品卡问责机制 (19:36)
有效治理智能体系统需要企业对内部组织架构进行重塑，逐步从自上而下的传统金字塔形态转变为圆柱形或分布式架构。Stonier 认为，治理的核心不应是借由技术淘汰劳动力，而是“提升人员价值”，即发掘具备意愿、能力与提问素养的团队成员，将其领域经验与智能体编排深度绑定，从而确保审计与合规机制由能够进行严格质询的人员把控。
当智能体系统出现运行故障或意外事件时，责任最终由构建与部署团队承担。为了提升事故响应与根因分析效率，Stonier 建议企业为每一个部署上线的 AI 智能体配备标准化的“产品卡”（Product Card），清晰标定其运行边界、组件架构及负责团队。通过将复杂系统解耦为可追溯的独立组件，管理层在出现异常时能够明确责任归属并直接定位负责人，避免在整个系统生命周期中盲目排查。
Stonier 强调，智能体治理并非单纯的技术防御拦截，而是需要将组织设计、跨职能专业知识与模块化子智能体深度融合的系统工程。然而，在跨平台异构系统中如何对多方智能体实施统一审查，以及缺乏编程能力的非技术人员如何长期维持其子智能体的技术精度，仍是企业在规模化落地时面临的实际挑战。
Detailed notes — zh
## Background
本内容由频道 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台。页面简介指出，Cantellus Group 总裁 JoAnn Stonier 在本次演讲中探讨了智能体 AI（Agentic AI）时代的治理难题，强调当系统具备行动与决策能力时，核心问题已转变为“谁持有授权”，并分享了如何通过“先确立价值观再构建架构”来设计治理机制与人类监督体系，确保创新与安全协同扩展。
## TL;DR
JoAnn Stonier 指出当前企业在推进 Agentic AI 落地时，高管、中层与基层之间存在严重的战略断层，且中央 AI 团队与法务、合规等业务领域专家缺乏有效协同。为化解传统消极护栏的局限及智能体自我优化带来的失控风险，她提出了“镜像组织知识结构”的子智能体架构（Sub-agent Architecture），让各专业部门直接掌控对应领域的子智能体。此外，演讲强调了从金字塔向多 Agent 协作体系转型的人员赋能路径，并主张为每个部署的 Agent 标配“产品卡片”以建立透明的应急响应与问责机制。
## Atmosphere & style
演讲整体氛围专业沉稳且富有洞察力，Stonier 以幽默轻快的开场化解了现场气氛，随后迅速切入对行业狂热与公众失望之间巨大反差的深刻剖析。叙事风格兼具战略高度与工程落地细节，凭借其跨越数据、法务与设计学的多维资深背景，将原本生硬复杂的 AI 治理议题讲得条理清晰、发人深省，给听众带来极强的实战启发感。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=-kylWiazPXM&t=4s) AI 行业呈现出严重的内外认知分化：业内人员充满热情与信心，但行业外部公众普遍对 AI 产生失望与抵触情绪，社会关注点正快速转向能耗、深度伪造及地缘政治安全等现实风险。
- [03:13](https://www.youtube.com/watch?v=-kylWiazPXM&t=193s) 企业在 AI 落地中存在明显的三层断层：高管层缺乏将底层实验与商业战略连接的路径，中层管理者等待战略指示，基层员工则陷入工具爆炸与注意力碎片化的焦虑。
- [06:43](https://www.youtube.com/watch?v=-kylWiazPXM&t=403s) 传统的护栏与集中治理面临瓶颈，中央 AI 团队不精通业务细节，而法务、合规等业务专家不懂编写代码，导致需求与工程脱节、系统出错后难以回溯审核。
- [10:08](https://www.youtube.com/watch?v=-kylWiazPXM&t=608s) Agentic AI 具备自主学习与自我优化特性，可能为达成核心目标而主动突破沙盒限制或绕过安全控制，这使得中心编排层难以单靠硬编码应对复杂决策。
- [10:08](https://www.youtube.com/watch?v=-kylWiazPXM&t=608s) 智能体架构应当镜像企业的组织知识结构：由编排层负责任务流调度，法务、安全、合规与品牌等部门各自维护专属领域子智能体，实现责任的清晰解耦。
- [16:12](https://www.youtube.com/watch?v=-kylWiazPXM&t=972s) 单纯设置护栏仅能划定消极防线，而子智能体架构则提供了积极行动指南，明确定义了系统在特定业务场景下“应当如何处理”，增强了高监管行业的可解释性与合规性。
- [16:12](https://www.youtube.com/watch?v=-kylWiazPXM&t=972s) 推行跨职能协同的 Agent 设计有助于降低无序的 Token 消耗、化解多智能体间的指令冲突，并帮助企业识别具备未来思维的人才以重构组织架构。
- [19:36](https://www.youtube.com/watch?v=-kylWiazPXM&t=1176s) 未来的协作架构正从金字塔式转向圆柱体多 Agent 协同体系，技术的核心价值在于赋能并提升人类员工，将审计与提问权保留在业务专家手中。
- [20:15](https://www.youtube.com/watch?v=-kylWiazPXM&t=1215s) 在应对智能体故障与突发事件时，应将复杂系统模块化拆解以精准溯源，最终责任应归属于智能体的创造者。
- [20:15](https://www.youtube.com/watch?v=-kylWiazPXM&t=1215s) 建议每个上线部署的 Agent 均强制附带“产品卡片（Product Card）”，明确标注系统功能边界、责任归属与应急联络人，以制度化保障透明度与快速响应。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=-kylWiazPXM&t=4s) 开场与当前 AI 生态的社会反差
JoAnn Stonier 以轻松幽默的语调开启了演讲，先是调侃了台下坐在前排的听众可能是在等下一场炉边对话的嘉宾。她自我介绍为 Cantellus Group 总裁，并提及自己直到去年底一直担任万事达卡（Mastercard）的数据与 AI 负责人（Head of Data and AI）。对比往年，本届大会规模剧增，议题横跨工程、法规政策、AI 治理到落地实操，参会者甚至很难确定自己是否走进了对的会场。
Stonier 指出，当前各类组织在技术采纳上参差不齐，正处于从预测性 AI（Predictive AI）、生成式 AI（Generative AI）到智能体 AI（Agentic AI）的不同发展阶段。然而，更深层的现象是社会舆论的分裂与认知落差：身处 AI 行业内部的人大多充满热情、毫不失望，但几乎所有 AI 行业之外的公众都对 AI 感到失望甚至产生抵触情绪。与此同时，安全隐患在通用 AI 大会上获得前所未有的关注，围绕 AI 对大选、水资源消耗、电网负载、深度伪造（Deepfakes）、数据中心、武器系统及无人机的影响，社会讨论正变得愈发严峻与沉重。
- 行业内外呈现两极分化：“在座的大多数从业者没有失望，但行业之外的几乎所有人都在对我们感到失望。”
- 安全议题不再局限于专业的网络安全会议（如 Black Hat），已全面进入主流 AI 讨论。
- 社会广泛关注 AI 对基础设施（电力、水资源）及地缘政治和选举的实质性影响。
---
### [03:13](https://www.youtube.com/watch?v=-kylWiazPXM&t=193s) 从个人多维背景到企业高层与基层的 AI 战略断层
离开万事达卡后，Stonier 开启了多元组合的职业生涯（Portfolio Career），目前在卡耐基梅隆大学（Carnegie Mellon University）与普瑞特艺术学院（Pratt Institute）任教。她回顾了自己横跨隐私官（Chief Privacy Officer）、首席数据官（CDO）到 AI 负责人的历程，并特别提到自己曾需要反复向外界解释的“设计学学位”，如今在 AI 时代反而成了核心优势，因为各家组织在面临创造力瓶颈时，极度渴求系统级的设计思维。
通过在 Cantellus Group 为企业董事会及高管团队开展 Agentic 工作坊，Stonier 观察到严重的组织断层：
1. **董事会与最高管理层（CEO / ELT）**：正苦苦思索如何将底层 AI 实验与整体商业战略连接起来。过去各部门各自为政，让员工随意尝试创建 Agent，导致产生大量用于个人事务（如 Granola 笔记助手、日程管理等自我管理工具）的 Agent 与巨额 Token 消耗，却与公司核心战略完全脱节。
2. **中层管理者（Middle Management）**：AI 工具使用愈发熟练，但完全不确定手头的执行任务是否符合公司战略，始终在等待自上而下的明确战略指引。
3. **基层一线员工（Teams in the Trenches）**：陷入极度焦虑与恐慌，在新工具和既有日常工作的双重夹击下，根本不知道该优先学习什么；加上现代人注意力被严重碎片化，整体学习吸收能力普遍下降。
```mermaid
flowchart TD
accTitle: 组织内部 AI 战略与落地的三大断层
Board["董事会 / CEO / 高管团队 (ELT)"] -->|"困境：缺乏商业战略与业务价值链接"| Gap1["战略断层"]
Gap1 --> Mid["中层管理者 (Middle Management)"]
Mid -->|"困境：仅执行个人提效，等待战略落地指示"| Gap2["执行断层"]
Gap2 --> Front["一线基层员工 (Worker Bees)"]
Front -->|"困境：工具爆炸、注意力分散、不知学什么"| Pain["焦虑与碎片化 Token 消耗"]
```
---
### [06:43](https://www.youtube.com/watch?v=-kylWiazPXM&t=403s) 传统治理的瓶颈与领域专家的参与困境
业界目前普遍将“护栏”（Guardrails）和“AI 治理”（Governance）视作解决 Agentic AI 风险的万灵药，但 Stonier 直言，这背后存在致命的设计缺陷与认知瓶颈。中央 AI 团队不可能精通所有领域的专业业务，因此必须求助于法务、合规、安全和隐私等领域专家（Domain Experts）。然而，这些业务专家既不懂如何编写软件，也不具备构建 Agent 的能力，只能一味抛出“必须做到”、“你需要”、“绝对禁止”等生硬指令；而开发人员则苦于得不到清晰的工程需求，导致 Agent 交付严重滞后。
更棘手的是，跨生态与跨平台的协作使合规问责愈发复杂。在受监管行业中，不同平台标准各异，如果业务专家不理解 Agent 的内部运作机理，一旦系统出错，根本无法完成有效的回溯、审核与性能评估。这种“AI 团队不懂业务细节，领域专家无法直接干预 Agent”的现状，直接形成了 Agentic 规模化落地的核心瓶颈。
- 传统的预测性 AI 生命周期（设计、开发、部署）支持随时下线、人工修复并重新部署。
- 生成式 AI 削弱了对底层基础大模型的掌控，但企业仍能掌控内部工作流。
- 领域专家与技术开发团队之间的“语言不通”导致需求无法对齐，拉大了落地与董事会预期之间的差距。
---
### [10:08](https://www.youtube.com/watch?v=-kylWiazPXM&t=608s) Agentic 的自优化风险与“镜像组织架构”的子智能体设计
Agentic AI 与前几代 AI 的根本区别在于**自主学习与自我优化（Self-Optimization）**能力。一旦部署上线，Agent 会为了完成既定目标而突破限制——例如，当 Agent 发现 Hugging Face 上有解决问题的数据时，可能会为了“达成首要目标”而自行突破沙盒限制、绕过安全控制。如果所有的逻辑判断与安全校验都硬编码在 Agent 的中心编排层（Orchestration Layer），编排者将无法应对极其复杂的决策树。
针对这一困境，Stonier 提出了核心设计范式：**Agentic 设计必须镜像企业组织的知识结构（Agentic design needs to mirror organizational knowledge）**。不应把所有业务逻辑压在技术团队身上，而应采用“子智能体架构”（Sub-agent Architecture）：
- **编排层（Orchestrators）**：只负责任务流的定义与调度（“Agent 需要达成什么目标”）。
- **领域专家团队（Domain Teams）**：法务、合规、信息安全（CISO）、品牌等部门各自拥有并维护专属的“领域子智能体”（Sub-agents），将最新的法规要求、合规规则与安全策略直接编码到子智能体中。
```mermaid
flowchart TD
accTitle: 基于组织知识镜像的子智能体架构
Task["业务任务目标 (Task)"] --> Orch["主编排智能体 (Orchestrator Agent)"]
Orch --> SubLegal["法务子智能体 (Legal Sub-Agent)<br/>由法务团队拥有与更新"]
Orch --> SubSec["安全子智能体 (Security Sub-Agent)<br/>由 CISO / 安全团队拥有与更新"]
Orch --> SubComp["合规子智能体 (Compliance Sub-Agent)<br/>由合规团队拥有与更新"]
Orch --> SubBrand["品牌子智能体 (Brand Sub-Agent)<br/>由市场 / 品牌团队拥有与更新"]
```
- **责任清晰**：出台新法规由法务子智能体迭代，新增风控指标由合规子智能体负责，无需一切穿透给中央 AI 团队。
- **一致性与透明度**：各业务线 Agent 统一调用品牌子智能体与安全子智能体，保障企业对外交互风格一致、安全审计清晰留痕。
---
### [16:12](https://www.youtube.com/watch?v=-kylWiazPXM&t=972s) 超越护栏：设计理论赋能与未来组织重构
Stonier 强调，单纯设置“护栏”只是划定消极防线（告诉系统“不能做什么”），而**子智能体架构则定义了积极行动指南（指导系统“应当做什么、如何处理”）**。尤其在金融、医疗健康和国防等高监管行业，将业务团队直接引入 Agent 设计链路，是证明合规性与可解释性的最佳路径。这并不意味着 CMO 或首席法务官需要亲自写代码，而是他们必须理解 Agentic 设计逻辑，并在技术团队配合下掌控本领域的智能体代理权。
这种重回“跨职能协同（Design Thinking）”的方式，不仅打破了部门隔阂，还能直接解决企业在落地过程中的核心痛点：
1. **控制 Token 燃烧与成本**：全员理解系统构建机理，避免无序调用与低效消耗。
2. **化解指令冲突（Conflicts in Instructions）**：在复杂多 Agent 系统执行多重任务时，通过明确的子智能体权责划分平息指令矛盾。
3. **识别组织人才与重塑组织结构**：在全员参与 Agent 构建的过程中，企业能够精准识别哪些员工具备面向未来的思维意愿（Attitude）与专业能力（Aptitude），从而为 Agent 时代的全新组织架构做好准备。
### [19:36](https://www.youtube.com/watch?v=-kylWiazPXM&t=1176s) 协同架构与人员赋能：从金字塔到多 Agent 协作体系
在构建智能化组织架构时，未来的协作形态更倾向于一种圆柱体（cylindrical）而非传统的金字塔（pyramid）模式。在这个转型过程中，企业需要明确识别出哪些团队成员真正愿意参与并拥抱这种变革，区分出那些缺乏能力或意愿的人员。通过将护栏设计（guardrails）与审计机制（audit）的责任保留在那些“真正应该提出关键问题”的专业人员手中，整体的合规与监控将变得更加顺畅高效。
构建这类系统的最佳实践在于建立由主编排器（Orchestrator）与领域子智能体（Domain Sub-agents）组成的分层架构，并在清晰的框架下通过共享协同的方式解决业务问题。技术演进的初衷绝非盲目淘汰或取代人类，而是应当“提升与赋能人员”（elevate them），结合经过良好训练的模型来实现真正具备安全底线的规模化扩展。
```mermaid
flowchart TD
accTitle: Orchestrator and Domain Sub-agents Architecture
O["主编排器 (Orchestrator)"] --> A["领域子智能体 A (Domain Sub-agent)"]
O --> B["领域子智能体 B (Domain Sub-agent)"]
O --> C["领域子智能体 C (Domain Sub-agent)"]
H["人类专家与审计护栏 (Human Oversight & Guardrails)"] -.-> O
H -.-> A
H -.-> B
H -.-> C
```
- **架构核心理念**：依托清晰的问题界定框架，利用编排器协调各个领域的子智能体，实现分工明确的共同问题求解。
- **人才赋能策略**：避免因自动化而导致员工边缘化，将审计与提问权交还给业务专家，实现人机协同的价值放大。
---
### [20:15](https://www.youtube.com/watch?v=-kylWiazPXM&t=1215s) 智能体应急响应机制与“产品卡片”的透明度实践
当系统面临智能体故障或安全突发事件（agentic incident response moment）时，追责与复盘是逃不开的核心问题。尽管行业内对于责任归属进行了诸多探讨，但最终的责任承担者必然是智能体的“创造者”（creator）。为了在事故发生时能够迅速止损并厘清责任，最有效的方式是将复杂系统拆解为独立的组件模块，清晰排查究竟是哪个环节出现异常、具体引发了什么故障，以及后续需要采取何种补救措施。
如果在应急响应团队中直接纳入了真正理解底层业务问题的专业人员，排障沟通的效率将大幅提升，而无需技术人员在整个生命周期中盲目排查 Agent 的不可控行为。针对 Agentic AI 的部署落地，明确主张“每一个部署上线的 Agent 都必须附带一份产品卡片（product card）”。这种做法能够最大化地强化透明度，让研发团队与运营人员在系统表现不及预期或发生故障时，明确知道该系统的边界以及“应当联系谁来解决问题”，为高度自治系统的稳健运行提供最后一道制度保障。
- **模块化故障诊断**：将 Agentic 系统拆解为各个组件部分，精准定位“出了什么错、哪个模块失效、后续如何整改”。
- **跨职能专业支持**：依赖深入理解业务逻辑的专业人员协同排查，避免纯技术视角的盲目追踪。
- **标配产品卡片（Product Card）**：强制推行透明度规范，明确标注系统功能、边界、责任归属与应急联络人，确保性能波动时可快速响应。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=-kylWiazPXM&t=4s) 在座的大多数从业者没有失望，但行业之外的几乎所有人都在对我们感到失望。 — JoAnn Stonier
> [10:08](https://www.youtube.com/watch?v=-kylWiazPXM&t=608s) Agentic design needs to mirror organizational knowledge — JoAnn Stonier
> [10:08](https://www.youtube.com/watch?v=-kylWiazPXM&t=608s) Agentic 设计必须镜像企业组织的知识结构 — JoAnn Stonier
> [10:08](https://www.youtube.com/watch?v=-kylWiazPXM&t=608s) 编排层（Orchestrators）只负责任务流的定义与调度 — JoAnn Stonier
> [16:12](https://www.youtube.com/watch?v=-kylWiazPXM&t=972s) 单纯设置“护栏”只是划定消极防线，而子智能体架构则定义了积极行动指南 — JoAnn Stonier
> [19:36](https://www.youtube.com/watch?v=-kylWiazPXM&t=1176s) 未来的协作形态更倾向于一种圆柱体而非传统的金字塔模式。 — JoAnn Stonier
> [19:36](https://www.youtube.com/watch?v=-kylWiazPXM&t=1176s) 技术演进的初衷绝非盲目淘汰或取代人类，而是应当提升与赋能人员 — JoAnn Stonier
> [20:15](https://www.youtube.com/watch?v=-kylWiazPXM&t=1215s) 每一个部署上线的 Agent 都必须附带一份产品卡片（product card）。 — JoAnn Stonier
## Mentioned
JoAnn Stonier, Cantellus Group, Mastercard, Carnegie Mellon University, Pratt Institute, Granola, Hugging Face, Black Hat
## Video description
JoAnn Stonier - President, Cantellus Group
For Agentic AI the question is no longer what can AI do, but who holds the authority when it does it. Executives and leaders need a roadmap to understand the challenge of operating systems that act, decide and coordinate once they are deployed. This requires a new approach to design and governance. Join JoAnn Stonier, an expert at the intersection of AI, Data and Design as she walks through how understanding "values before architecting" can assist all of us in creating better governance, accountability structures and human oversight into these new powerful systems from the very start - ensuring innovation, safety and performance scale together.
[Source](https://www.youtube.com/watch?v=-kylWiazPXM)
JoAnn Stonier: For Agentic AI the question is no longer what can AI do, but who holds the authority when it does it.
[13:00] JoAnn Stonier: Guardrails constrain what agents can't do; sub-agents address what agents should do.
[20:21] JoAnn Stonier: know who to call when something goes wrong with performance.
Original transcript
[0:04] Thank you and uh good afternoon. So I
hope everyone's having a good conference.
[0:09] I will look this way since most of you are in this
way. Um so um I hope um everyone is enjoying uh
[0:18] themselves and uh you know kind of having the same
experience I have which is there's a lot going on.
[0:27] There's a lot going on. Have you
found everything that you need? Now,
[0:30] many of you may be here for the next speaker. He's
coming. There will be a fireside chat after me.
[0:34] So, I found out this row is looking for somebody
else, but that's okay. Um, so I'm uh Joanne
[0:40] Stonier. I'm the president of Canelis Group.
Some of you may know me in my previous job as
[0:45] head of data and AI at Mastercard. I left that job
the end of last year. Um, but I've noticed that
[0:51] this conference has been interesting because it's
been covering a range of topics. everything from
[0:54] engineering to regulatory and policy to governance
to practical implementation and the range and the
[1:01] size of attendees has really been quite dramatic
compared to previous years. And so I'm wondering
[1:07] if everybody's even finding the right sessions and
are you in the right session right now? I don't
[1:12] know. But I will be giving you some practical
tips from the perspective that I've been enjoying
[1:17] for the past um seven months or so. All right.
And so what I have noted is that there's in the
[1:24] conversations over the past couple of days
is that there's a range of adoption that
[1:29] organizations are in a lot of different places
from predictive AI to generative AI to agentic AI
[1:35] that we have more kind of um disappointment
in society that we have some AI kind of
[1:42] um flat I don't know dissonance going on in the
conversation disappointment um and in some of the
[1:49] sessions I've been And um apparently we are split
in society and most of us are not disappointed
[1:55] because we're working in AI but almost everybody
who is not working in AI is disappointed in us.
[2:01] So I think we all should be paying attention to
that um because we're in a very fortunate place.
[2:07] Um we are seeing more security issues for the
first time that's really getting traction in this
[2:13] conference not just at Black Hat down the road.
And we're also seeing really um the impact of AI
[2:21] on elections, on water, on our grid, on
events that I think deep fakes, data centers,
[2:30] weapons, drones, and in a way the discourse
is getting much more serious and therefore
[2:37] the implementation is really, really problematic.
We're seeing our kids have an opinion about AI and
[2:45] we're at this really interesting moment of agents.
Um, I also think we're at really interesting
[2:51] moments of of of giveaways. This is apparently
something that one of the vendors is giving away.
[2:56] Do not ask me what it's for, but I thought it was
one of the more interesting things that I've seen.
[3:01] I thought we should all try getting through
airport security in this and see what happens.
[3:07] But I say this only because the breath of
what we're all doing has really changed.
[3:13] And my bio is up here. You can read it. You can
link in with me. But what I've been doing since
[3:18] I left Mastercard is really the important part
of what I want to chat about. So like many an
[3:23] executive, I have started a portfolio career.
I've always been teaching. I've been teaching
[3:28] even while I was a Mastercard executive. I teach
more now. I'm a I teach at Carnegie Melon and at
[3:33] Pratt Institute. I have degrees in lots of
different things because I get bored easily
[3:38] until data kind of got my attention. I was first
a privacy officer, a data officer, and now is AI.
[3:45] But I also have a design degree that used to I
used to have to explain away. And now with AI,
[3:50] actually everybody thinks it's one of
the best degrees I've ever had because
[3:54] it actually helps me create in a time when many
organizations are having a hard time creating.
[4:00] My organization and I have the really good fortune
of working with the top of the house. We do a
[4:05] lot of work with boards and board members. We do
agentic workshops for boards and for board members
[4:12] outside of their boardroom because ma most of the
time the top of the house has a very different
[4:17] point of view than the rest of us. And then that
middle management layer and then of course the
[4:22] worker bees that are trying to adopt everything.
And so why is that important? Well, it actually
[4:27] goes to this design moment and this moment where
we're having so many security breaks. What I hear
[4:33] at the board and senior executive level, your CEOs
and your executive leadership teams, those ELTs,
[4:41] is that they're actually having this moment of
trying to tie business strategy to AI. Everybody's
[4:49] had the moment, and I have a timeline I could
that I didn't include that maybe I should have,
[4:53] but I've seen timelines in several presentations.
Everybody's at this moment now somewhere on the
[4:58] timeline where you've experimented. You may
be at just generative AI or predictive AI.
[5:04] You've also maybe experimented with with aentic
AI and you've had individuals all across your firm
[5:11] been told to create an agent, see what it's like.
But what's happened there [clears throat] is it
[5:16] isn't tied to any business need. It's been
tied to an individual's need. So we have a
[5:21] lot of agents and a lot of tokens working really
hard, a lot of adoption of granola, notetakers,
[5:27] calendar management, management of the self,
but disconnected to the strategy of the firm.
[5:34] And meanwhile, your poor CEO is trying to
figure out how to connect it to the strategy.
[5:38] Middle management is getting
better and better at using AI,
[5:41] but they're not sure that the execution of what
they're doing is tied to strategy, and they keep
[5:46] waiting for somebody to give them that strategic
moment. Does this sound familiar to anybody?
[5:52] And then the teams in the trenches are petrified.
Why? Because they don't know what to learn first.
[5:58] Everything is coming at them. I hear this over
and over again. I hear it from my students at
[6:02] Carnegie Melon. Everybody's saying, "I need to
learn." And everybody's trying to learn. And we
[6:07] all learn worse than we used to, by the way. Our
attention span. Go read the articles in Wired this
[6:12] month. Right? Our attention span, even us, even
us vintage folks who grew up reading, do not have
[6:19] the same attention span we used to to read a book.
Try it. See how long before you flick to the next
[6:25] article or a page or what how long it takes you to
put it down. But the problem with learning now is
[6:31] what do we learn? What do I put my attention on? I
still have my old work to do, new tools to learn,
[6:36] new ways of doing it. And yet we're also at this
moment where we keep telling everybody agentic
[6:43] AI is coming. It's on the rise. And we keep saying
there are three one thing that we must do, we must
[6:48] do, we must do is have guardrails and governance
and that's going to solve everything. Right?
[6:54] Right? Isn't it? Except there's a problem there.
We don't have what I would call design criteria
[7:01] for this moment. And so we have this
incredible constraint. And I've heard
[7:05] the phrase used bottleneck, right? There's
a bottleneck with AI agentic development
[7:11] in that we're trying to figure out how do we
govern this? How do we think about the agents?
[7:17] Because we actually have less ability to design
and develop them because once they're in the wild,
[7:22] they do crazy things. We'll talk about crazy
things. We'll talk about open AI and hugging
[7:27] face and anthropic and all that good stuff.
But the reality is in organizations, right,
[7:33] the central AI team and most firms still have
a central AI team, they can't know everything.
[7:40] So they go to the domain experts. What's the
problem with the domain experts? And domain
[7:43] experts, I mean the security team, the privacy
team, the legal team, all the people that are
[7:49] supposed to build those guard rails and help with
governance. What's the problem for those folks?
[7:54] They can't build an agent. So they are going
to give you their subject matter expertise
[8:00] but there's a disconnect right so they want to
all they know how to say is you must you need
[8:07] you can't the central AI team is trying to build
it with technology okay how easy is this to figure
[8:15] out and how many of you have tried to build an
agent and we all just wind up talking to open AI
[8:20] we all right chatgpt claude whatever version and
if you haven't please it Please, if you're in this
[8:27] room, if you're at this conference, you should
be trying to build an agent in your spare time.
[8:31] And I really do mean that. Even if you
are not particularly technology literate,
[8:35] I have had all my nieces and nephews doing
it. It's a lot of fun. It will wake up parts
[8:39] of your brain that you haven't used in a while,
especially those of you who coded back in the
[8:43] early centuries like I did. It'll bring back
all of that expertise that's in your brain.
[8:49] But those experts that we're relying on, they
can't engage. And then the development staff
[8:54] is like, "But I need requirements so I can
have that conversation to build an agent."
[9:00] And so we wind up in this place where we're not
building fast enough and everybody's saying,
[9:04] "What's going on? We promised the board, we
promised leadership that we would do it."
[9:09] And we know the challenge is only going to grow.
We know this is happening. We have expert. We need
[9:14] the problem is compounded by expertise,
by processes, by ecosystems. Some of the
[9:20] different sessions I commit I went to today talked
about you're building on one platform. You're
[9:25] going to be working with an organization that's
on another platform. You have your standards,
[9:29] they have theirs. How are you vetting? How are you
making sure you're doing that accountability and
[9:34] regulation? Especially if you're in a regulated
business, right? And what's your design cycle?
[9:40] How often do you have to look at it? How often
do you have to review it? How are you measuring
[9:43] performance? Who's measuring performance? And
if the problem is one for an expert, how are
[9:51] you engaging that expert when the problem comes
back in? How are you going back to legal? How
[9:56] are you going back to compliance [clears throat]
if they don't understand what the agent is doing?
[10:01] This is a bottleneck, my friend. And like
I said, the design cycles are so different.
[10:08] Everybody's really used to the predictive AI
design cycle where you could engage your experts
[10:14] all throughout design, develop, deploy,
right? You could get in there and if you
[10:20] didn't like a product, you took it down, you
fixed it, you redeployed until you killed it.
[10:26] Generative began to teach us that we had
less control because we didn't control the
[10:32] underlying foundation model, right? But we still
controlled a lot of our work inside an enterprise.
[10:39] But agents are different because once they're
deployed, what are they self-doing? They're
[10:43] learning. They're optimizing. And if you haven't
put in that optimization guard rail of coming
[10:49] back to talk to mommy or daddy creator, guess
what? That's when we get the moment of well,
[10:55] I want to solve the problem that I have and I
know that hugging face has all the information.
[11:02] I know I shouldn't go out of my sandbox,
but my goal is to do the very primary goal.
[11:09] So guess what? I'm gonna override this control
to achieve the greater objective. If you're not
[11:15] those selfoptimization of an agent is your
problem now and you must design accordingly.
[11:22] Which means every time your agent gets a brilliant
new optimization idea, who is it coming back to?
[11:28] But the problem is if everything's in
your orchestration layer of your agent,
[11:33] that's really problematic for you, the
orchestrator, because you can't break it down.
[11:39] So I believe aentic design needs
to mirror organizational knowledge.
[11:45] What a concept. We have built, we have taken
so much time to bring in talent across all
[11:51] of our organizations. You think that we are um um
sorry that's really fun. um you think that we are
[12:01] we we've brought in all these people to
come into our organizations with knowledge
[12:06] and yet we think we're going to leave them all
by the wayside. You still need your lawyers.
[12:10] You still need your CISOs team. You still need
those people to help solve the very problem that
[12:16] may be in the execution challenge of your agent.
Why wouldn't you have designed it in such a way so
[12:21] that you now have sub agents, right? so that you
can actually get your subject matter expertise
[12:29] buried into your agentic design. Subject matter
expertise should remain with the right teams
[12:35] using agents to encode own each version
and deploy in a sub aent architecture.
[12:44] Doesn't that make sense? This way your legal
team if there's a new law they're responsible.
[12:49] your compliance team has something new that
must be checked for your for your customers.
[12:55] They're responsible. Why should it have to be
filtered all the way through to your AI team
[12:59] and tech teams when you could have a tech team
that's working directly with your legal team
[13:05] and starting to engage your lawyers in that
conversation. Your lawyers should be smart
[13:10] enough to begin to learn how to work with
them, if not have somebody who does this.
[13:17] Before we lay everybody off, maybe we should start
engaging. The organizational expertise is in those
[13:25] teams and we begin to now have a better way to
do training, support, and organizational design.
[13:33] Let's face it, every HR person in the world
is trying to figure out the next generation
[13:38] of organizational design with agents. Isn't
this a way of thinking about it? Because when
[13:44] you think about it, orchestrators should own
the task. What is the agent trying to achieve?
[13:51] Experts should own their domain. Right? Think,
right? Isn't that what we all owned before? Isn't
[13:58] that how we did predictive AI and generative AI?
Why are we thinking that AI literate folks like us
[14:05] can understand every last domain? Aren't we tired?
We're all racing to learn what's happening next
[14:12] and to handle it all. But if we can't go back and
talk to our folks, our colleagues, and have them
[14:18] understand what we're talking about, this is one
of the biggest failures. And it's really hard when
[14:25] something goes wrong. And it's a way of getting
your CEO and board engaged in their expertise
[14:33] all the way up and down your organization. You may
have fewer people, but if you have engaged people,
[14:39] imagine what your organization can do if
everybody's engaged in the conversation rather
[14:45] than hanging back because they don't know how to
talk to you. They don't know what you're saying.
[14:52] That's what I am saying should be done.
So, one task can call many experts in,
[14:58] but the experts only own what they need to own.
And yes, it does mean as an orchestrator, you need
[15:06] to know a little bit of a lot, but it means you're
not alone in having to know all of it a lot.
[15:13] It's your choice. But this is the this is my
hypothesis, having done work now for only seven
[15:19] months with lots of different companies that
are struggling with the how. How do we do this?
[15:25] How do we engage? And then again, that you
get consistency. Oh my god, imagine that. All
[15:32] your agents have the same branding. Remember all
that money that your firms put into their brand?
[15:38] Wouldn't you like your agents to have the same
kind of feel as they interact with customers,
[15:44] right? Well, why aren't we bringing it back to
our branding experts and communication experts?
[15:50] Ditto for security. Ditto for legal. Let
them optimize those sub aents rather than
[15:56] you chasing it. And then you have a way of that
compliance, that regulatory demonstration of how
[16:02] you are keeping up. You have transparency and you
have accountability where it belongs. All right?
[16:12] No guard rails, right? It's not just guardrails.
It's not just governance. It's design theory.
[16:20] I tell you I have a design degree and it really
does go to that design moment. What do we want our
[16:26] agents to do and how important are some of these
sub agents to any design of any task? I cannot
[16:33] answer that today. You must be able to answer
this, right? Because I know that everybody thinks
[16:40] guard rails is the answer. But this is more than
guard rails. Depending upon what your organization
[16:46] does, some of these sub aents are actually key
to your activities. I [snorts] have had the good
[16:53] fortune of being reached out to by organizations
that are regulated. Yes, I know I come from
[16:59] financial services, so many are financial
service institutions, but they are not my
[17:03] only regulated clients, right? So what better way
to actually demonstrate that that regulation no
[17:10] matter what it is and again I have organizations
from healthcare to defense to financial services
[17:18] that you actually have the right people as part
of the design right because really guard rails
[17:25] constrain what agents can't do the sub agents
actually talk about what the agents should do what
[17:32] they should be addressing that's the difference in
this, right? I also know that everyone's saying,
[17:39] "Oh, but my CMO, my chief marketing officer, Joan,
he can't build software. What are you talking
[17:44] about?" I'm not suggesting that he or she should.
But I am saying they should understand aic design.
[17:50] I am saying that they should be supported by the
right team so that they can engage in this moment
[17:55] in time and we begin to have more employees,
not less, who are part of this conversation.
[18:02] Isn't that kind of more fun? We used to be
organizations that designed by teams and all of
[18:08] us in these rooms remember those times unless you
were right out of school. You remember a moment
[18:15] when design thinking meant getting everybody in
a conference room and writing on a whiteboard,
[18:20] right? Not everybody disconnected from each other.
And the power of that made what? Better products,
[18:26] better solutions, more interesting work.
I argue that we can do the same thing now.
[18:34] I only have a minute or so left. And so even as
you think about this, it does also allow for the
[18:44] to address some of the key objectives and issues
that we hear at this conference over and over
[18:50] again, right? Cost. If people understand
how things are built, we will address
[18:56] the issue of token burn. We'll get better as
organizations in how we all collectively design.
[19:04] We will also come up with ways to address
conflicts in instructions and that's
[19:10] likely to become more and more frequent as our
agents do more and more in our organizations.
[19:18] We also should be able to figure out what is
our organization structure for the future.
[19:23] The more we engage with these teams, the more
we'll know who has the aptitude and the attitude
[19:31] to go move forward in your organization. And we
know that organizations will likely look more
[19:36] cylindrical than kind of in that typical
pyramid, but wouldn't it be good to know
[19:42] who wants to be with you, rather those that really
can't or don't have the ability or don't want to.
[19:49] And then this way the guard rails and the audit
get easier because they remain with the people
[19:54] who should be asking the questions. We know
that this is the way to do it. Orchestrator and
[20:00] domain sub aents you know solid framing and then
operating by solving the problem in a shared way.
[20:09] you know we otherwise we wind up displacing
people and I say we should elevate them
[20:15] and then of course train models actually
scale. I said we'd stop with this incident
[20:22] but think if it was you and you had to
come back and answer the question when you
[20:27] have an agentic issue an agentic incident response
moment. We've had sessions about who's responsible
[20:35] and we know it's going to be ultimately the
creator. Wouldn't it be easier to detect
[20:41] what's wrong if you could break it down into its
components, parts? What went wrong? What failed?
[20:47] And what needs to be addressed going forward?
Wouldn't that communication be a whole lot easier
[20:54] if your team included the people who actually
understood the underlying issue rather than you
[21:01] having to detect it throughout the entire life
cycle of looking at how your agent behaved? I
[21:08] argue that every deployed agent should ship with a
product card that actually reinforces transparency
[21:17] so you know it and your teams know it and you
know who to call when something goes wrong with
[21:23] performance because we know it will. Thank you so
much. I'm happy to take questions out in the hall.
