Kenvue AI Head Urges Small Models to Avert Surging Agent Token Costs

Extracting Value from Agentic AI - A Pragmatic Approach — Ai4

Ai4

Enterprise scaling of agentic AI requires shifting from commercial flagship LLMs to fine-tuned small models and standardized protocols like MCP to avoid unsustainable token and infrastructure expenses.

AI-assisted summary of Extracting Value from Agentic AI - A Pragmatic Approach — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 24, 2025, this presentation features Rahul Tyagi, VP & Global Head of Data Science, AI & Intelligent Automation at Kenvue. According to the page description, the session provides practical, ready-to-execute strategies to help organizations navigate the roadmap toward scaling agentic AI adoption and realizing measurable business value.

## TL;DR Rahul Tyagi delivers an enterprise roadmap for adopting agentic AI, tracing the evolution from legacy rule-based automation to modern learning-driven architectures. He details the foundational layers of agent creation, orchestration, and model selection, while highlighting the critical need for balanced governance and evaluation frameworks. Tyagi also explains multi-agent interoperability using protocols like Google's A2A and the Model Context Protocol (MCP), concluding with FinOps strategies that leverage fine-tuned Small Language Models (SLMs) and quantization to keep token costs and Total Cost of Ownership sustainable.

## Atmosphere & style The presentation feels clear, authoritative, and grounded, striking an effective balance between high-level executive strategy and technical architecture. Tyagi begins with an interactive, conversational icebreaker about household consumer brands before shifting into a measured, analytical breakdown of agentic systems. The delivery cuts through contemporary AI hype, creating a realistic and practical atmosphere focused on operational trade-offs, engineering discipline, and economic sustainability.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=xnp4zk04zXU&t=4s) An AI agent is characterized by agency—the autonomous capacity to pursue specific goals and make decisions within strict policy boundaries and contextual prompts. - [00:35](https://www.youtube.com/watch?v=xnp4zk04zXU&t=35s) Agentic systems represent an evolution from deterministic, rule-based systems like RPA and expert systems to learning-driven models powered by LLMs and reinforcement learning. - [00:35](https://www.youtube.com/watch?v=xnp4zk04zXU&t=35s) An enterprise agentic architecture requires three fundamental tiers: the Application/Creation Layer, the Orchestration Layer, and the Model Bedrock. - [06:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=360s) Agent implementation requires continuous balancing of the trade-off triad: operational latency, execution cost, and output accuracy. - [06:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=360s) Because agent evaluation standards are still maturing, enterprises currently rely on hybrid validation frameworks combining human-in-the-loop review and 'LLM as a judge' methods. - [07:39](https://www.youtube.com/watch?v=xnp4zk04zXU&t=459s) Interoperability protocols like Agent-to-Agent (A2A) enable cross-communication between disparate frameworks such as Azure, LangGraph, and CrewAI. - [07:39](https://www.youtube.com/watch?v=xnp4zk04zXU&t=459s) The Model Context Protocol (MCP) provides a standardized abstraction layer between agents and external tools, separating the tool catalog (MCP Client) from execution (MCP Server). - [07:39](https://www.youtube.com/watch?v=xnp4zk04zXU&t=459s) Direct custom APIs are preferable for connecting to a minimal number of endpoints, while MCP should be leveraged when scaling across large tool ecosystems. - [13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) Unmonitored token consumption from multi-agent iterations and MCP tool listings serves as the primary ongoing cost driver in agentic deployments. - [13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) Long-term FinOps and TCO optimization require replacing costly flagship commercial LLMs with domain-specific Small Language Models (SLMs) using PEFT and quantization.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=xnp4zk04zXU&t=4s) Introduction and Defining AI Agents

Rahul Tyagi opens with a quick interactive quiz, asking how many audience members know what Kenvue is. Finding very few hands raised, he pivots to familiar household brands: Listerine, Aveeno, Neutrogena, Johnson's Baby, Tylenol, and Motrin. He clarifies that Kenvue is the company owning these iconic consumer healthcare brands, having spun off from Johnson & Johnson Consumer a couple of years prior. Setting the stage for his talk on extracting value from Agentic AI, he promises to provide enough technical clarity so leaders can ask the right questions of their technology teams, platform providers, and service partners, ensuring a future-proof foundation despite rapid industry changes.

He anchors the discussion on the core definition of an AI agent, highlighting "agency" as a fundamental willfulness to carry out acts based on goals, autonomous decision-making, and policy boundaries. To illustrate, he uses the scenario of planning a one-week family trip to Italy for four people within a $5,000 budget. While the agent operates autonomously, it requires detailed instructions via prompts—such as specific destinations like Naples, Rome, and Venice—to deliver the expected results. Furthermore, the agent operates within a "policy framework," which in business terms translates to operational rules, such as accounts receivable and accounts payable terms in a corporate free cash flow agent.

- **Kenvue Portfolio Brands:** Listerine, Aveeno, Neutrogena, Johnson's Baby, Tylenol, Motrin. - **Key Characteristics of AI Agents:** - Clear goal or objective (e.g., $5,000 budget, 7-day Italy trip for a family of four). - Autonomous action without continuous human supervision. - Policy adherence specifying constraints, operational bounds, and detailed context.

[00:35](https://www.youtube.com/watch?v=xnp4zk04zXU&t=35s) Evolution from Rule-Based Systems to Learning Agents

The discussion traces the historical evolution of agents, stressing that agentic concepts are not entirely new. Rule-based agents have existed for years within Robotic Process Automation (RPA) and older expert systems utilizing forward-chaining rules. What distinguishes modern Agentic AI is the shift toward learning-driven systems powered by AI and techniques like reinforcement learning, which are used to train both language models and complex decision-making agents.

```mermaid flowchart LR accTitle: Evolution of Agentic Systems subgraph Legacy["Rule-Based Systems"] A["Expert Systems / Forward Chaining"] --> B["Robotic Process Automation (RPA)"] end subgraph Modern["AI-Driven Agents"] C["Language Models (LLMs)"] --> D["Reinforcement Learning Agents"] end Legacy --> Modern ```

To establish an enterprise-wide agentic foundation, organizations must address three foundational layers: 1. **Application / Creation Layer:** Determining whether agent creation is low-code/no-code for widespread use or centralized within a specialized team. 2. **Orchestration Layer:** Managing inter-agent communication, external data sources, and cross-framework interoperability (e.g., Azure, CrewAI, LangGraph, Google). 3. **Model Bedrock:** Choosing between commercial proprietary large language models and customized, fine-tuned models.

[06:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=360s) Agent Creation, Evaluation, and Governance

Designing effective agents requires balancing autonomy, operational guidelines, and rigorous evaluation mechanisms. Because language models naturally risk hallucination, robust evaluation frameworks are essential. Rahul notes that current agent evaluation techniques are still maturing and predicts major advancements in standardized, robust evaluation frameworks by the first half of next year. Currently, enterprise teams rely on hybrid approaches, combining human judgment, "LLM as a judge," and customized internal knowledge alongside emerging external benchmarks.

- **Autonomy vs. Control:** Establishing hybrid control models with overarching operational guidelines. - **Evaluation Approaches:** Utilizing human-in-the-loop validation alongside "LLM as a judge" frameworks. - **Multi-Agent Coordination:** Managing conflicting sub-agent goals and reconciling multi-agent outputs. - **Trade-off Triad:** Constantly balancing execution cost, accuracy, and operational latency.

[07:39](https://www.youtube.com/watch?v=xnp4zk04zXU&t=459s) Technical Architecture: Interoperability and MCP Integration

Navigating multi-agent environments often involves distinct, coexisting frameworks—such as Microsoft Azure alongside open-source alternatives like LangGraph or CrewAI. To allow agents across different ecosystems to communicate, frameworks like Google's Agent-to-Agent (A2A) protocol serve as abstraction layers. However, Rahul cautions against unconstrained proliferation, as excessive inter-agent connectivity increases management overhead and infrastructure costs. Enterprise leaders must balance the simplicity of limiting standard frameworks against the flexibility of allowing teams to select their preferred tools.

```mermaid flowchart TD accTitle: Enterprise Agentic System Architecture subgraph Frameworks["Agentic Frameworks"] A1["Azure Agent Framework"] <-->|A2A Protocol| A2["Open Source Framework"] end subgraph Orchestration["Abstraction & LLM"] A1 & A2 --> B["Model Context Protocol (MCP) Client"] B --> C["Language Model Bedrock"] end subgraph Execution["Resource Integration"] B --> D["MCP Server"] D --> E1["Travel / Booking APIs"] D --> E2["Enterprise Databases"] D --> E3["External Websites"] end ```

The Model Context Protocol (MCP) provides a standardized abstraction layer connecting agentic frameworks to external resources, databases, and third-party web services. Rather than building custom direct APIs for every integration point, MCP simplifies access across extensive tool ecosystems. Within the architecture, the **MCP Client** maintains an inventory of accessible tools and resources, while the **MCP Server** functions as the execution engine, dispatching requests, filling forms, and executing external actions.

- **A2A Protocol:** Facilitates abstraction and communication across heterogeneous agentic frameworks. - **MCP Client:** Functions as a catalog listing accessible internal databases, travel portals, or financial systems. - **MCP Server:** Acts as the execution engine that queries APIs, fetches data, and carries out autonomous transactions. - **Integration Rule:** Use direct custom APIs when connecting to very few resources; leverage MCP when scaling across numerous endpoints.

[13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) Model Strategy and FinOps Optimization

The presentation concludes with a focus on financial and architectural sustainability. Language model token usage represents the largest ongoing cost component in any agentic framework, as every query, tool lookup, and step consumes prompt and response tokens. The Model Context Protocol, if unmonitored, can consume massive quantities of tokens through resource listings and repeated tool requests.

To optimize the three-year Total Cost of Ownership (TCO) and establish strong FinOps practices, Rahul advises against relying solely on general-purpose commercial flagship LLMs. Instead, organizations should deploy smaller, distilled, or fine-tuned language models (SLMs). Techniques such as Parameter-Efficient Fine-Tuning (PEFT) and model quantization deliver high accuracy and fast execution at a fraction of the compute cost.

- **Cost Management:** Carefully monitor token consumption across MCP requests and system prompts. - **Model Efficiency:** Implement Parameter-Efficient Fine-Tuning (PEFT) and quantization to optimize performance. - **Architecture Choice:** Prefer domain-focused, fine-tuned Small Language Models (SLMs) over heavy commercial models for routine task execution.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=xnp4zk04zXU&t=4s) agency is the willfulness to carry out acts based on goals, autonomous decision-making, and policy boundaries. — Rahul Tyagi

> [00:04](https://www.youtube.com/watch?v=xnp4zk04zXU&t=4s) planning a one-week family trip to Italy for four people within a $5,000 budget. — Rahul Tyagi

> [00:35](https://www.youtube.com/watch?v=xnp4zk04zXU&t=35s) What distinguishes modern Agentic AI is the shift toward learning-driven systems powered by AI and techniques like reinforcement learning — Rahul Tyagi

> [06:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=360s) agent evaluation techniques are still maturing and predicts major advancements in standardized, robust evaluation frameworks by the first half of next year. — Rahul Tyagi

> [06:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=360s) utilizing human-in-the-loop validation alongside 'LLM as a judge' frameworks. — Rahul Tyagi

> [07:39](https://www.youtube.com/watch?v=xnp4zk04zXU&t=459s) excessive inter-agent connectivity increases management overhead and infrastructure costs. — Rahul Tyagi

> [07:39](https://www.youtube.com/watch?v=xnp4zk04zXU&t=459s) The MCP Client functions as a catalog listing accessible internal databases, travel portals, or financial systems. — Rahul Tyagi

> [07:39](https://www.youtube.com/watch?v=xnp4zk04zXU&t=459s) The MCP Server functions as the execution engine that queries APIs, fetches data, and carries out autonomous transactions. — Rahul Tyagi

> [07:39](https://www.youtube.com/watch?v=xnp4zk04zXU&t=459s) Use direct custom APIs when connecting to very few resources; leverage MCP when scaling across numerous endpoints. — Rahul Tyagi

> [13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) Language model token usage represents the largest ongoing cost component in any agentic framework — Rahul Tyagi

> [13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) deploy smaller, distilled, or fine-tuned language models (SLMs). Techniques such as Parameter-Efficient Fine-Tuning (PEFT) and model quantization deliver high accuracy and fast execution at a fraction of the compute cost. — Rahul Tyagi

## Mentioned Rahul Tyagi, Kenvue, Johnson & Johnson Consumer, Listerine, Aveeno, Neutrogena, Johnson's Baby, Tylenol, Motrin, Microsoft Azure, CrewAI, LangGraph, Google, Agent-to-Agent (A2A) Protocol, Model Context Protocol (MCP), Robotic Process Automation (RPA), Parameter-Efficient Fine-Tuning (PEFT)

## Video description Rahul Tyagi, VP & Global Head (Data Science, AI & Intelligent Automation), Kenvue

Abstracts

Is your organization ready to deploy Agentic Ai at scale? While the value potential of Agentic AI is clear, most organizations lack a clear and pragmatic roadmap of how to embed it in various functions. Agentic AI requires a close partnership with business functions to deliver the promised value. Business functions need a very clear value articulation and realization plan to be convinced. In this session, you will learn practical and ready-to-execute tips to accelerate you journey towards value realization through scaling adoption of agentic AI in your organization.

[Source](https://www.youtube.com/watch?v=xnp4zk04zXU)

中文简报

Kenvue AI主管呼吁采用小模型以应对智能体Token成本激增 企业规模化落地智能体AI需从商用旗舰LLM转向微调小模型及MCP等标准化协议，以避免难以承受的Token与基础设施开销。 Rahul Tyagi提出了企业落地智能体AI的路线图，梳理了从传统基于规则的自动化向现代学习驱动型架构的演进历程。他详细阐述了智能体构建、编排与模型选择的基础架构层，并强调了构建平衡的治理与评估框架的关键必要性。Tyagi还解析了基于Google A2A与模型上下文协议（MCP）等多智能体互操作方案，最后给出了利用微调小语言模型（SLM）与量化技术的FinOps策略，以确保Token成本与总体拥有成本维持在可持续水平。 AI Agent的核心在于“自主能动性”（Agency），能够在严格的策略边界与提示词约束下自主决策以达成特定目标，正从传统的RPA与专家系统向由LLM及强化学习驱动的学习型系统演进。 (0:04) 企业级Agent架构包含应用/创建层、编排层与模型基座三层结构，在落地时必须持续平衡操作延迟、执行成本与输出准确性这三者之间的权衡。 (0:35) 针对评估标准尚不成熟的现状，企业目前主要结合人工在环验证与“LLM as a judge”方法，并借助A2A等互操作协议实现Azure、LangGraph与CrewAI等异构框架间的跨系统通信。 (6:00) 模型上下文协议（MCP）通过MCP Client与MCP Server分离工具目录与执行操作，适合大规模工具生态的扩展，而少数端点连接则更适合直接构建定制API。 (7:39) 为控制Token消耗并优化三年的总体拥有成本（TCO），企业应引入FinOps实践，结合量化与PEFT等技术，优先采用经过微调的小型语言模型（SLM）替代通用旗舰LLM。 (13:01)

Reading summary

Rahul Tyagi of Kenvue argues that extracting sustainable enterprise value from agentic AI requires moving beyond basic rules to autonomous learning systems bounded by strict operational policies, standardized integration layers, and cost-conscious model architectures. Defining Autonomous Agents within Policy Frameworks (0:04) Modern agentic systems mark an evolution from legacy rule-based automation, such as Robotic Process Automation (RPA) and forward-chaining expert systems, toward learning-driven models trained with reinforcement learning. Tyagi defines an AI agent by its agency: the capacity to autonomously pursue a designated objective without continuous human oversight, operating within explicit policy constraints. To succeed, agents require structured operational bounds alongside detailed prompt instructions. For example, an agent tasked with planning a seven-day family vacation to Italy for four within a $5,000 budget requires specific destination prompts for cities like Rome and Venice. In a corporate environment, this translates to policy boundaries such as accounts receivable and payable terms governing a free cash flow agent. Managing Governance, Evaluation, and Framework Interoperability (6:00) Enterprise agent deployments must continuously balance a triad of operational trade-offs: execution cost, latency, and accuracy, particularly given the inherent risk of large language model hallucinations. Because agent evaluation methodologies are still maturing, Tyagi notes that organizations currently rely on hybrid solutions blending human-in-the-loop oversight, internal domain knowledge, and 'LLM as a judge' techniques. He forecasts that standardized, robust evaluation frameworks will see major advancements by the first half of next year. Governance also extends to orchestrating diverse multi-agent frameworks, such as reconciling Microsoft Azure environments with open-source tools like LangGraph or CrewAI. While abstraction protocols such as Google's Agent-to-Agent (A2A) facilitate cross-framework communication, Tyagi warns against unconstrained connectivity, which risks significant management overhead and escalating infrastructure costs. Scalable System Integration with MCP (7:39) Connecting agents to external data sources, enterprise databases, and web services requires clear architectural boundaries. Tyagi advises utilizing direct, custom APIs when integrating with very few endpoints, but recommends adopting the Model Context Protocol (MCP) as an abstraction layer when scaling integrations across extensive tool ecosystems. Within this architecture, the MCP Client functions as a central registry cataloging accessible resources and tools, while the MCP Server serves as the execution engine. The server handles operational requests, queries external databases, inputs forms, and performs autonomous transactions across third-party endpoints without requiring bespoke point-to-point connections. Mitigating Token Costs Through Model Optimization (13:01) Ongoing token consumption represents the primary cost driver in agentic systems, as multi-step autonomous workflows, detailed system prompts, and repetitive MCP tool listings rapidly deplete compute budgets. Unmonitored, tool lookups and resource queries can compromise the financial viability of an agent deployment. To optimize the three-year Total Cost of Ownership (TCO) and enforce FinOps discipline, Tyagi advises against defaulting to general-purpose commercial flagship models for everyday tasks. Instead, organizations should deploy smaller, domain-distilled language models (SLMs) and apply techniques such as Parameter-Efficient Fine-Tuning (PEFT) and model quantization to maintain operational accuracy and low latency at a fraction of the cost. Realizing enterprise value from agentic AI requires balancing autonomy with rigorous policy enforcement and architectural discipline. As organizations wait for evaluation standards to mature, long-term viability will depend on curbing framework sprawl and containing token expenses through smaller, fine-tuned models.

精炼阅读摘要

Kenvue技术发言人Rahul Tyagi指出，企业若要从智能体AI（Agentic AI）中获取商业价值，必须建立清晰的分层架构与政策边界治理，并在自主性、准确度与运行成本之间寻求平衡。他强调，控制以Token为主的持续成本与规范跨框架工具集成，是决定智能体能否规模化落地的核心前提。 智能体核心定义与三层基础架构 (0:04) Rahul Tyagi将AI智能体的核心特征定义为“能动性”（Agency），即智能体依据明确的目标和政策框架，自主进行决策并执行任务。例如，在规划人均预算5000美元、四人一周的意大利旅行时，智能体虽自主运行，但仍依赖详细提示词限定具体目的地；在企业场景中，政策框架则体现为自由现金流智能体所遵循的应收和应付款项业务规则。从技术演化来看，智能体并非新概念，其前身为机器人流程自动化（RPA）与专家系统等基于前向链规则的架构，而现代智能体则转向由大语言模型和强化学习驱动的学习型系统。 构建企业级智能体基础需要确立三个层级：一是应用与创建层，决定智能体开发是采用面向全员的低代码/无代码工具，还是集中由专业团队构建；二是编排层，负责管理智能体之间的通信、外部数据集成以及Azure、CrewAI、LangGraph与Google等多生态间的互操作；三是模型底层，需在商用专重大语言模型与定制微调模型之间做出架构取舍。 治理边界、协同冲突与评估成熟度 (6:00) 设计高效智能体需要在赋予其自主权与施加操作指引之间建立混合控制模式。鉴于语言模型天然存在幻觉风险，严谨的评估机制不可或缺。Tyagi表示，目前的智能体评估技术仍在成熟过程中，企业现阶段主要采取混合评估策略，即结合人工介入（Human-in-the-loop）验证、“大语言模型作为裁判”（LLM as a judge）框架、内部知识库与新兴外部基准。根据他的预测，行业将在明年上半年迎来重大的标准化、健壮的评估框架突破。 随着架构向多智能体协同演进，系统必须妥善处理子智能体之间的目标冲突，并有效调和多智能体的输出差异。工程团队在部署时必须持续权衡“成本、准确度与延迟”这三者构成的铁三角，避免因单方面追求智能体的无监督自治而导致执行延迟超标或计算资源浪费。 跨框架互操作性与模型上下文协议 (7:39) 在企业多智能体环境中，往往并存着微软Azure及LangGraph、CrewAI等开源平台。为实现跨生态通信，系统可借助Google的智能体间通信协议（A2A）等抽象层完成对接；但Tyagi提醒，无节制的智能体互联会激增管理负担和基础设施成本，管理者需在标准化单一框架的简洁性与多工具选择的灵活性间取得平衡。 对于智能体访问外部资源与数据库的需求，模型上下文协议（MCP）提供了标准化的抽象支持。在架构中，MCP Client负责维护内部数据库、订票门户或财务系统等工具的清单目录；MCP Server则作为执行引擎，承担调用API、提取数据及执行事务操作。针对集成路线，Tyagi提出了明确准则：当系统仅需连接极少数端点时，开发定制的直接API更为高效；只有在需要横向扩展并对接大量外部服务时，采用MCP才能体现出架构优势。 Token成本控制与小模型微调战略 (13:01) 智能体系统在工程层面的长期生命力取决于财务运维（FinOps）的优化。在智能体运行过程中，语言模型的Token消耗是最大的持续成本项，智能体分解任务、检索工具及中间交互均会持续消耗提示词与回复Token。若缺乏监控，MCP协议在列出工具资源和反复调用工具时容易产生巨大的Token开销。 为了优化三年期的总体拥有成本（TCO），Tyagi建议企业避免在所有环节单纯依赖昂贵的通用商用旗舰LLM。企业应转向部署参数量较小、经过蒸馏或领域微调的小语言模型（SLM）。通过运用参数高效微调（PEFT）和模型量化等技术，企业能够以极低的算力成本换取高任务准确率与低执行延迟，从而实现智能体架构的经济可持续性。 Rahul Tyagi的论述表明，企业从智能体AI中提取实际价值的关键，在于确立严格的政策边界治理、规范工具与数据接口抽象，并通过微调小模型有效抑制Token消耗的增长；但多智能体评估标准与互操作生态目前仍处于演化初期，需要工程团队在实践中持续权衡成本与灵活性。

Detailed notes — zh

## Background 该视频由 Ai4 于 2025 年 8 月 24 日发布在 YouTube 平台上，记录了 Kenvue 全球数据科学、AI 与智能自动化副总裁 Rahul Tyagi 的演讲。视频简介指出，演讲聚焦于企业如何制定务实的 Agentic AI 落地路线图，通过业务与技术协同加速规模化部署与价值变现。 ## TL;DR Rahul Tyagi 系统阐述了企业规模化落地 Agentic AI 的三层架构：涵盖评估机制与自主权衡的智能体创建层、支持多框架协同与 MCP 协议的编排层，以及决定长期总体拥有成本（TCO）的模型底座层。他重点剖析了模型上下文协议（MCP）在异构资源调用中的架构机制与严重的 Token 消耗风险，并对比了专用 API 与 MCP 的适用边界。最后，他提出了借助小语言模型（SLM）、模型蒸馏与量化等精细化模型工程来控制推理成本的方法，旨在帮助业务与技术管理者建立底层技术认知并提出关键架构问题。 ## Atmosphere & style 整场分享节奏极快且信息密度极高，展现出典型的大型跨国企业资深技术主管沉稳、务实且直击要害的演讲风格。讲者以家喻户晓的消费品牌与通俗的旅行规划案例作为引子，随后迅速切入到协议规范、通信开销与算力经济学等硬核架构剖析中。整体氛围冷静而专业，既客观肯定了前沿技术的潜力，又清醒指出了工具评估缺失、Token 消耗黑洞及架构蔓延等现实陷阱，极具实战指导意义。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=xnp4zk04zXU&t=4s) 理解 Agentic AI 底层逻辑的核心目的，是帮助业务与技术领导者在面对技术团队或外部方案商时能够提出正确的技术问题，确立稳固的关键技术锚点。 - [01:30](https://www.youtube.com/watch?v=xnp4zk04zXU&t=90s) 现代 AI Agent 具备三大核心支柱：明确的任务目标（Goal）、自主拆解执行能力（Autonomy）以及必须严格遵守的业务策略框架（Policy）。 - [03:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=180s) 现代智能体与传统规则型 Agent（如 RPA 和专家系统）的本质区别，在于引入了具备强化学习能力的现代语言模型，具备自主学习与适应能力。 - [04:36](https://www.youtube.com/watch?v=xnp4zk04zXU&t=276s) 企业构建 Agentic AI 基础设施需要系统性布局三层架构：智能体创建层（Agent Creation）、编排协同层（Orchestration）以及底层语言模型层（Foundation Models）。 - [06:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=360s) 由于底层模型存在固有幻觉，构建智能体评估框架（结合 LLM as a Judge 与人工评估）是确保业务落地的关键前提，目前多数企业仍需自研补齐这一能力。 - [07:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=420s) 在编排层面对 CrewAI、LangGraph 等多元框架共存时，可借助 Google A2A 等协议互通，但企业需在组件选型自由度与多框架治理成本之间取得平衡。 - [08:46](https://www.youtube.com/watch?v=xnp4zk04zXU&t=526s) 模型上下文协议（MCP）作为连接智能体与海量外部资源的标准化抽象层，适用于大规模异构系统对接；若仅连接少数固定系统，直接使用原生 API 调用更轻量高效。 - [10:30](https://www.youtube.com/watch?v=xnp4zk04zXU&t=630s) MCP 架构由维护可用资源黄页清单的 MCP Client 与负责具体任务执行调用的 MCP Server 构成，企业需根据业务职能划分独立的 MCP 拓扑结构。 - [12:00](https://www.youtube.com/watch?v=xnp4zk04zXU&t=720s) MCP 默认机制极易引发“Token 消耗风暴（Token Guzzling）”，高频交互与资源列表维护会导致成本急剧上升，必须进行严格的 Prompt 优化与监控。 - [13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) 在 Agentic AI 的三年 TCO 与 FinOps 评估中，底层模型层往往是最大的开支来源，盲目依赖通用顶配大模型容易导致成本失控。 - [13:30](https://www.youtube.com/watch?v=xnp4zk04zXU&t=810s) 降低智能体运行成本的最佳实践包括：结合业务场景采用模型微调与蒸馏、选用小语言模型（SLM）、应用参数高效微调（PEFT）以及模型量化压缩技术。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=xnp4zk04zXU&t=4s) 引言与 Kenvue 品牌背景 Rahul Tyagi 在开场时先向观众做了一个快速互动小测验，询问现场有多少人知道 Kenvue（科赴）这家公司。看到举手的人寥寥无几后，他接着列举了几个家喻户晓的消费品牌：Listerine（李施德林）、Aveeno（艾惟诺）、Neutrogena（露得清）、Johnson's Baby（强生婴儿）、Tylenol（泰诺）以及 Motrin（美林）。他解释道，Kenvue 原本是强生公司（Johnson & Johnson）旗下的消费健康业务板块，在两年前被独立分拆出来；虽然公司新名字大家可能还不熟悉，但几乎所有人都曾使用过旗下的某款产品。 随后他切入演讲的核心议题：如何从 Agentic AI（代理式人工智能）中真正获取业务价值。他提到前一位演讲者 Venkat 的内容非常出色，深入探讨了 Agentic AI 成功的关键要素。而本场分享的侧重点在于提供一定深度的技术认知，帮助业务与管理人员掌握底层逻辑，从而能够向内部技术团队或外部解决方案与服务提供商“提出正确的技术问题”。由于该领域技术迭代极快，企业必须提前确立关键技术锚点并打好地基，才能在技术不断演进的过程中保持架构的稳固与敏捷。 ### [01:30](https://www.youtube.com/watch?v=xnp4zk04zXU&t=90s) 什么是 AI Agent：核心特征与业务策略约束 要理解 Agentic AI，首先必须回归本质。AI Agent（智能体）的核心在于具备“Agency”（意图性或主动执行特定行动的意愿）。一个合格的智能体通常具备以下几个关键特征： 1. **目标或任务导向（Goal/Objective）：** 例如输入一个明确任务——“为一家四口规划为期一周的意大利行程，预算为 5000 美元”。 2. **自主行动能力（Autonomy）：** 智能体能够在无人监督的情况下自主拆解并解决问题，例如计算 7 天时间、4 人出行、5000 美元预算下的最优方案。但为了让结果符合预期，人类需要通过提示词（Prompt）提供详尽的细节（如指定游览那不勒斯、罗马与威尼斯，以及各城市的停留天数和特定偏好景点）。 3. **策略框架约束（Policy）：** 智能体必须在既定的策略边界内运作。在个人场景中，策略体现为预算上限、旅行地点及停留时间；在企业业务场景中，策略则体现为具体的运营规范。例如在财务领域的自由现金流（Free Cash Flow）管理中，应收账款（AR）与应付账款（AP）的贸易结算条款，就是现金流管理 Agent 必须严格遵循的“Policy”。 ```mermaid flowchart LR accTitle: AI Agent 核心运作机制 subgraph Input["输入层"] A["明确目标<br/>(Goal)"] B["上下文提示<br/>(Prompt)"] end subgraph Core["智能体核心"] C["自主决策与学习<br/>(AI Agent)"] D["业务策略边界<br/>(Policy Framework)"] end subgraph Output["行动层"] E["自动化执行结果<br/>(Action / Execution)"] end A --> C B --> C D -. 约束与规范.-> C C --> E ``` Agent 本身并非全新概念。早期的规则驱动型智能体（Rule-based Agents）已存在多年，最典型的是 RPA（机器人流程自动化）以及更早期的专家系统（Expert Systems），它们依靠前向链式规则（Forward Chaining Rules）按照预设层级逻辑执行任务。如今的根本变革在于引入了具备学习能力、由 AI 驱动的现代智能体。强化学习（Reinforcement Learning）已成为训练大语言模型和智能体行为的核心技术。 ### [04:36](https://www.youtube.com/watch?v=xnp4zk04zXU&t=276s) 组织构建 Agentic AI 的三层基础设施 企业在搭建面向未来的 Agentic AI 技术底座时，需要系统性地考量三个关键层级：应用与创建层、编排层以及底座语言模型层。 ```mermaid flowchart TD accTitle: Agentic AI 组织级三层架构 subgraph AppLayer["1. 应用与创建层 (Agent Creation Layer)"] A1["自动化 / 低代码 / 无代码构建"] A2["评估机制 (LLM / Human as a Judge)"] A3["自主程度与业务规则权衡"] end subgraph OrchLayer["2. 编排与协同层 (Orchestration Layer)"] B1["多智能体协作框架 (CrewAI / LangGraph / 云原生)"] B2["跨框架通信协议 (Google A2A)"] B3["工具与资源抽象 (MCP 协议)"] end subgraph ModelLayer["3. 底座模型层 (Foundation Language Models)"] C1["商业闭源模型 vs 开源专有模型"] C2["模型微调 / 蒸馏 (Fine-tuning / Distillation)"] C3["小语言模型与量化 (SLM / Quantization)"] end AppLayer --> OrchLayer OrchLayer --> ModelLayer ``` - **智能体创建层（Agent Creation Layer）：** 企业需要决策智能体的构建模式——是采用自动化、低代码/无代码平台供全员使用，还是设立专业团队集中承接需求并开发？同时要明确智能体的自主程度（是否构建混合型 Agent，在自主决策的同时受到顶层规则约束）、训练成本控制（平衡成本、准确性与延迟），以及至关重要的**评估框架（Evaluation Framework）**。由于底层依赖的模型存在固有幻觉（Hallucination），必须引入评估机制（如“LLM as a Judge”、“Human as a Judge”或两者结合）。目前的评估工具仍不完善，多数先行企业只能自行补充开发评估能力，预计更成熟的框架要到次年上半年才会逐渐普及。 - **智能体编排层（Agent Orchestration Layer）：** 关注智能体之间、以及智能体与外部数据和网站等资源的交互。目前市面上存在大量的开源与专有框架（如微软 Azure、AWS、Google 云平台方案，以及开源的 CrewAI、LangGraph 等）。当企业内部共存多个框架时，如何实现互通成为难题。Google 推出的 A2A（Agent-to-Agent）协议提供了跨框架通信的标准抽象；但如果企业内部任由不同框架泛滥，A2A 连接带来的管理开销和系统成本也会激增。因此必须在“使用便利性”（各团队自由选型）与“架构简洁性”（限制框架数量、降低运维负担）之间取得平衡。 - **底层语言模型（Language Model Bedrock）：** 决定是直接调用第三方商业模型，还是基于自建与微调的专有模型，这是决定长期总体拥有成本（TCO）的基石。 ### [08:46](https://www.youtube.com/watch?v=xnp4zk04zXU&t=526s) 模型上下文协议（MCP）与外部资源调用 在多智能体系统与外部环境交互时，模型上下文协议（Model Context Protocol, MCP）扮演着极其关键的角色。MCP 是连接智能体编排系统与外部资源（如数据库、第三方网站或企业内部应用）之间的抽象层。以预订意大利机票为例，智能体可以通过 MCP 路由访问 Expedia，自动填写表单，甚至在被完全授权的情况下直接在预算内完成扣款出票。 针对“既然已经有 API，为什么还需要 MCP”的疑问，Tyagi 指出核心判断标准在于**系统需要对接的外部资源数量**： - 如果智能体只需要对接极少数固定的外部系统，直接编写 API 调用更为轻量高效，无需引入 MCP。 - 如果系统需要动态、大规模地访问海量异构数据源、外部网站或企业应用，MCP 提供的标准化抽象层将带来极大的扩展优势。 ```mermaid flowchart TD accTitle: MCP Client 与 MCP Server 协作机制 subgraph AgentSystem["智能体系统"] Agent["AI Agent / LLM"] end subgraph MCPStructure["模型上下文协议 (MCP)"] Client["MCP Client<br/>(维护可用资源列表清单)"] Server["MCP Server<br/>(负责具体任务的执行引擎)"] end subgraph ExternalResources["外部数据与服务资源"] R1["旅游/订票网站 (Expedia 等)"] R2["住宿平台 (Airbnb / VRBO)"] R3["本地游与活动数据"] R4["内部业务应用 / ERP / CRM"] end Agent -->|"发送查询请求"| Client Client -->|"识别所需资源"| Server Server -->|"调用与填写表单"| R1 Server -->|"匹配房源位置与价格"| R2 Server -->|"查询一日游路线"| R3 Server -->|"业务数据读写"| R4 Server -.->|"汇总结果返回"| Agent ``` MCP 内部主要由两个核心组件构成： - **MCP Client：** 相当于可用资源的“黄页清单”，记录了当前环境能够访问的所有数据源与服务平台（如机票网站、Airbnb/VRBO 等民宿平台、本地一日游服务等）。 - **MCP Server：** 实际的“执行引擎”。当接收到具体任务（如“在罗马停留 2 天，寻找合适住处”）时，MCP Server 会解析需求并向具体的服务发送调用请求，筛选出符合位置与预算的最佳选项，最终将结果汇总返回给 Agent。 在落地部署 MCP 时，企业必须重点关注几大关键问题： - **MCP 服务器拓扑结构：** 根据业务职能配置独立的专用服务器（如旅游、财务、营销等各设专属 MCP Server）。 - **Token 消耗风暴（Token Guzzling）：** MCP 默认机制极其消耗 Token。无论是在 MCP Client 端维护详尽的资源列表，还是 MCP Server 发起密集的资源请求，都会产生庞大的 Token 开销，这是 Agentic 架构中最昂贵的成本来源之一。因此必须极度严密地优化 Prompt 设计并全程监控请求量。 - **质量与输出评估：** 持续监控外部系统返回结果的准确度与服务质量。 ### [13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) 底层模型策略与成本控制（FinOps / TCO） 演讲最后，Tyagi 深入剖析了语言模型层的经济学考量。在评估为期三年的总体拥有成本（TCO）或采用 FinOps（云财务运营）框架时，模型层往往是整个 Agentic 架构中花销最大的地方，因为每一次交互、规划、工具调用和响应都在持续产生 Token 消耗。 为了防止成本失控，企业不应盲目依赖最昂贵的商业通用大模型，而应采取精细化的模型工程策略： - **模型蒸馏与微调（Distilled / Fine-tuned Models）：** 针对具体业务场景训练专有模型。 - **选用小语言模型（SLM）：** 多数特定业务任务并不需要超大参数模型，轻量化小模型完全能够胜任。 - **参数高效微调（PEFT / UFT）：** 降低微调阶段的算力消耗与工程复杂度。 - **模型量化（Quantization）：** 压缩模型体积，提升推理速度并显著降低部署门槛与运行成本。 Tyagi 总结道，尽管内容信息量巨大且节奏飞快，但理解这些架构逻辑与权衡考量，是每一位业务与技术领导者在面对技术团队与外部供应商时，问出关键问题、规避落地陷阱的核心底气。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=xnp4zk04zXU&t=4s) 本场分享的侧重点在于提供一定深度的技术认知，帮助业务与管理人员掌握底层逻辑，从而能够向内部技术团队或外部解决方案与服务提供商“提出正确的技术问题”。 — Rahul Tyagi > [00:04](https://www.youtube.com/watch?v=xnp4zk04zXU&t=4s) 由于该领域技术迭代极快，企业必须提前确立关键技术锚点并打好地基，才能在技术不断演进的过程中保持架构的稳固与敏捷。 — Rahul Tyagi > [01:30](https://www.youtube.com/watch?v=xnp4zk04zXU&t=90s) AI Agent（智能体）的核心在于具备“Agency”（意图性或主动执行特定行动的意愿）。 — Rahul Tyagi > [01:30](https://www.youtube.com/watch?v=xnp4zk04zXU&t=90s) 在财务领域的自由现金流（Free Cash Flow）管理中，应收账款（AR）与应付账款（AP）的贸易结算条款，就是现金流管理 Agent 必须严格遵循的“Policy”。 — Rahul Tyagi > [01:30](https://www.youtube.com/watch?v=xnp4zk04zXU&t=90s) 强化学习（Reinforcement Learning）已成为训练大语言模型和智能体行为的核心技术。 — Rahul Tyagi > [04:36](https://www.youtube.com/watch?v=xnp4zk04zXU&t=276s) 由于底层依赖的模型存在固有幻觉（Hallucination），必须引入评估机制（如“LLM as a Judge”、“Human as a Judge”或两者结合）。 — Rahul Tyagi > [04:36](https://www.youtube.com/watch?v=xnp4zk04zXU&t=276s) 必须在“使用便利性”（各团队自由选型）与“架构简洁性”（限制框架数量、降低运维负担）之间取得平衡。 — Rahul Tyagi > [08:46](https://www.youtube.com/watch?v=xnp4zk04zXU&t=526s) 如果系统需要动态、大规模地访问海量异构数据源、外部网站或企业应用，MCP 提供的标准化抽象层将带来极大的扩展优势。 — Rahul Tyagi > [08:46](https://www.youtube.com/watch?v=xnp4zk04zXU&t=526s) MCP 默认机制极其消耗 Token。无论是在 MCP Client 端维护详尽的资源列表，还是 MCP Server 发起密集的资源请求，都会产生庞大的 Token 开销，这是 Agentic 架构中最昂贵的成本来源之一。 — Rahul Tyagi > [13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) 在评估为期三年的总体拥有成本（TCO）或采用 FinOps（云财务运营）框架时，模型层往往是整个 Agentic 架构中花销最大的地方，因为每一次交互、规划、工具调用和响应都在持续产生 Token 消耗。 — Rahul Tyagi > [13:01](https://www.youtube.com/watch?v=xnp4zk04zXU&t=781s) 理解这些架构逻辑与权衡考量，是每一位业务与技术领导者在面对技术团队与外部供应商时，问出关键问题、规避落地陷阱的核心底气。 — Rahul Tyagi ## Mentioned Rahul Tyagi, Kenvue, Johnson & Johnson, Listerine, Aveeno, Neutrogena, Johnson's Baby, Tylenol, Motrin, Venkat, CrewAI, LangGraph, Microsoft Azure, AWS, Google, A2A (Agent-to-Agent), MCP (Model Context Protocol), Expedia, Airbnb, VRBO ## Video description Rahul Tyagi, VP & Global Head (Data Science, AI & Intelligent Automation), Kenvue Abstracts Is your organization ready to deploy Agentic Ai at scale? While the value potential of Agentic AI is clear, most organizations lack a clear and pragmatic roadmap of how to embed it in various functions. Agentic AI requires a close partnership with business functions to deliver the promised value. Business functions need a very clear value articulation and realization plan to be convinced. In this session, you will learn practical and ready-to-execute tips to accelerate you journey towards value realization through scaling adoption of agentic AI in your organization. [Source](https://www.youtube.com/watch?v=xnp4zk04zXU)

Source quotes

[0:04] Rahul Tyagi: 本场分享的侧重点在于提供一定深度的技术认知，帮助业务与管理人员掌握底层逻辑，从而能够向内部技术团队或外部解决方案与服务提供商“提出正确的技术问题”。 [0:04] Rahul Tyagi: 由于该领域技术迭代极快，企业必须提前确立关键技术锚点并打好地基，才能在技术不断演进的过程中保持架构的稳固与敏捷。 [1:30] Rahul Tyagi: AI Agent（智能体）的核心在于具备“Agency”（意图性或主动执行特定行动的意愿）。 [1:30] Rahul Tyagi: 在财务领域的自由现金流（Free Cash Flow）管理中，应收账款（AR）与应付账款（AP）的贸易结算条款，就是现金流管理 Agent 必须严格遵循的“Policy”。 [1:30] Rahul Tyagi: 强化学习（Reinforcement Learning）已成为训练大语言模型和智能体行为的核心技术。 [4:36] Rahul Tyagi: 由于底层依赖的模型存在固有幻觉（Hallucination），必须引入评估机制（如“LLM as a Judge”、“Human as a Judge”或两者结合）。 [4:36] Rahul Tyagi: 必须在“使用便利性”（各团队自由选型）与“架构简洁性”（限制框架数量、降低运维负担）之间取得平衡。 [8:46] Rahul Tyagi: 如果系统需要动态、大规模地访问海量异构数据源、外部网站或企业应用，MCP 提供的标准化抽象层将带来极大的扩展优势。 [8:46] Rahul Tyagi: MCP 默认机制极其消耗 Token。无论是在 MCP Client 端维护详尽的资源列表，还是 MCP Server 发起密集的资源请求，都会产生庞大的 Token 开销，这是 Agentic 架构中最昂贵的成本来源之一。 [13:01] Rahul Tyagi: 在评估为期三年的总体拥有成本（TCO）或采用 FinOps（云财务运营）框架时，模型层往往是整个 Agentic 架构中花销最大的地方，因为每一次交互、规划、工具调用和响应都在持续产生 Token 消耗。 [13:01] Rahul Tyagi: 理解这些架构逻辑与权衡考量，是每一位业务与技术领导者在面对技术团队与外部供应商时，问出关键问题、规避落地陷阱的核心底气。

Source quotes

[0:04] Rahul Tyagi: agency is the willfulness to carry out acts based on goals, autonomous decision-making, and policy boundaries. [0:04] Rahul Tyagi: planning a one-week family trip to Italy for four people within a $5,000 budget. [0:35] Rahul Tyagi: What distinguishes modern Agentic AI is the shift toward learning-driven systems powered by AI and techniques like reinforcement learning [6:00] Rahul Tyagi: agent evaluation techniques are still maturing and predicts major advancements in standardized, robust evaluation frameworks by the first half of next year. [6:00] Rahul Tyagi: utilizing human-in-the-loop validation alongside 'LLM as a judge' frameworks. [7:39] Rahul Tyagi: excessive inter-agent connectivity increases management overhead and infrastructure costs. [7:39] Rahul Tyagi: The MCP Client functions as a catalog listing accessible internal databases, travel portals, or financial systems. [7:39] Rahul Tyagi: The MCP Server functions as the execution engine that queries APIs, fetches data, and carries out autonomous transactions. [7:39] Rahul Tyagi: Use direct custom APIs when connecting to very few resources; leverage MCP when scaling across numerous endpoints. [13:01] Rahul Tyagi: Language model token usage represents the largest ongoing cost component in any agentic framework [13:01] Rahul Tyagi: deploy smaller, distilled, or fine-tuned language models (SLMs). Techniques such as Parameter-Efficient Fine-Tuning (PEFT) and model quantization deliver high accuracy and fast execution at a fraction of the compute cost.

Original transcript

[0:04] Firstly, uh just a quick quiz. Uh how [0:06] many of you how many of you know what [0:08] Canv is? [0:11] Not many. Okay. How many of you are [0:13] familiar with the following brands? Uh [0:16] Listen, [0:18] then u Aino, [0:21] Neutrogena, Johnson's Baby, Talenol, [0:25] Motrin. So these are the brands that are [0:27] owned and sold by by Canv. So this [0:30] company originally was a part of Johnson [0:31] and Johnson consumer got spun off a [0:33] couple of years ago. So the name is not [0:35] familiar but the brands I'm assuming all [0:37] of you uh probably have used one or one [0:39] or more of the brands at some point of [0:41] time. So today what I'm going to talk [0:44] about is uh how do we get value out of [0:46] Agentic AI. I think uh the previous [0:48] presentation by Vakat was a really good [0:50] presentation. uh you really got into [0:52] some some things that I think are [0:53] critical for success of aentic AI. Here [0:56] what I'll do is I'm going to get a bit [0:58] technical but again not a whole lot [1:00] technical. The key theme here is to come [1:03] up with an understanding where you can [1:05] ask the right technical questions of [1:08] your technology team or if you're [1:10] working with external partners solution [1:12] service platform providers you have the [1:14] you ask them the right technical [1:16] questions so that you have the right [1:17] foundation. There's a lot of stuff going [1:19] on in this this area. it's moving really [1:21] fast. So it's important to lay down a [1:23] certain kind of key stakes on the ground [1:26] so that even if technology evolves, you [1:28] have the right foundation in place. So [1:30] that's basically what the idea of this [1:31] presentation is. Now again, uh what's an [1:35] agent? I mean, let's get down to the [1:36] basics first. An agent here, an AI agent [1:39] specifically is is an agent that has an [1:41] agency. Agency is a willfulness to do a [1:44] certain act. So there are three or four [1:47] key characteristics when you talk about [1:48] an agent. One, there should be a goal or [1:51] objective. So, hey, plan a trip uh plan [1:55] a trip for a family of four to Italy for [1:57] a week. My budget is $5,000. So, that's [2:00] a goal. Now, it should be able to act [2:03] autonom autonomously without human [2:04] supervision. So, basically, it should be [2:07] able to figure out that okay, 7 days, [2:09] Italy, $5,000, family of four, how best [2:11] to do it. But that itinerary may not [2:14] match it. So one key thing that you have [2:16] to remember is you have to give a lot of [2:20] details to an agent to perform it the [2:23] way you want it to. So it needs to know [2:25] that you are going to visit you're going [2:26] to visit Naples, Rome and and and [2:29] Venice. It needs to know how many days [2:31] you want to do. It needs to know if you [2:34] have specific places that you are really [2:36] keen on doing. So that's something that [2:39] it that you will need to provide it [2:41] what's typically called in the form of a [2:43] prompt. [2:45] Now it acts within the confines of a [2:47] policy. So this is another word you will [2:49] hear as as agentic becomes more popular. [2:51] You'll hear of this thing called policy. [2:53] It's policy in this case that I [2:56] described is basically what you wanted [2:58] to do. This is my budget. These are the [2:59] places I want to go. These are the [3:01] number of days I want to spend in Rome. [3:02] So that's the policy framework within [3:05] which the agent will will work. [3:07] In a in a business context, this becomes [3:10] basically how it operates. So for [3:12] example, if you are talking about let's [3:14] say a finance function and you're [3:15] talking about uh free free cash flow uh [3:18] management and you're talking about [3:20] account receivable and account payable, [3:22] the the terms of trade for account [3:24] receivable per account payable would be [3:27] the policy in which that uh basically [3:30] free cash flow agent would operate. [3:35] So there's a whole so agents again [3:38] agents per se are not new. The [3:41] rule-based agents have been around for a [3:43] while. Uh traditionally in the fold of [3:45] uh this area called RPA. Even before [3:48] that there used to be an area called [3:49] expert systems where you used to build [3:51] rulebased systems what's called forward [3:52] chaining rules where you could chain up [3:54] rules and the agent would execute those [3:57] rules based upon the logic defined in [4:00] that hierarchy of rules. So they have [4:01] been around for a while. Uh RP has been [4:04] a is a very mature field. What's [4:06] changing now is what's this this AI this [4:09] learning this so-called I mean people [4:11] are calling it sensient still ways to go [4:14] but AIdriven learn learning agent which [4:17] learns based upon based upon some of the [4:19] policies and frameworks that I talked [4:20] about now again when we say learning [4:23] that's a whole separate area I'll cover [4:26] it a bit but it's it's a vast area uh [4:28] reinforcement learning is is a very [4:30] common technique that's used to train [4:32] both language models and and train the [4:34] agents [4:36] So if we look at if you're looking at [4:39] building let's say agentic foundation [4:42] for the for the organization there are [4:43] three key parts you need to think about. [4:45] How do I build the agent? So the top [4:47] layer what's what's called the [4:48] application layer the agent create [4:50] creation layer how should I create the [4:52] agent should be automated low code no [4:55] code or you want that to be limited to a [4:58] certain team so that that team receives [4:59] request to an agent and then creates it. [5:01] So those are the kind of decisions that [5:02] you need to take about agent creation, [5:04] agent orchestration. I think vocado uh [5:07] the CTO vocado covered it well. How does [5:11] this agent talk to other agents? How [5:13] does this agent talk to let's say other [5:15] resources like data resources like [5:17] external websites? [5:19] Maybe you want to use more than one [5:21] framework. So there are multiple [5:22] frameworks out there right now. They're [5:24] open source. They're proprietary [5:25] framework. All the hyperscalers [5:27] Microsoft AWS uh Azure have uh Azure and [5:31] and Google have their own frameworks lot [5:34] of open source crew langraph and and [5:36] many more are there. So maybe [5:38] organization has multiple frameworks. So [5:40] how do these different orchestration [5:41] frameworks talk to each other that's a [5:43] key consideration uh that you will need [5:45] to think of in this layer and then [5:47] fundamentally the kind of the bedrock on [5:49] which all of this stands the language [5:50] model you take a commercial model or you [5:53] basically build a proprietary model. So [5:55] these are the consideration. I'll get [5:56] into details of each of these in [5:58] subsequent slides. [6:00] So agent creation. So now here the one [6:04] of the key things you need to think [6:05] about is how much autonomy do you want [6:07] the agent to have? Do you want to have a [6:09] hybrid agent which has autonomy but [6:11] still there kind of a over overarching [6:13] set of rules or or guidelines within [6:15] which it operates. Then how do you train [6:18] it? Now training basically is is a [6:22] fairly intensive process and I'll talk a [6:23] bit more about when I talk about the [6:25] language models here. It's a it's a it's [6:27] a fairly costly process as well. You [6:28] need to balance cost, accuracy, latency [6:32] uh when you are thinking about these [6:33] kind of things. And very importantly, [6:36] how do you evaluate an agent? Since the [6:38] underlying foundation of agentic is a [6:41] language model, lagman models inherently [6:43] helps hallucinate. So having an [6:47] evaluation framework is extremely [6:49] critical. This you can have a human as a [6:52] judge, LLM as a judge or a combination [6:53] of the two. There are there are existing [6:56] form basically frameworks around agent [6:58] evaluation not sufficient. So this is [7:01] something that is probably going to [7:02] evolve. My guess is probably during the [7:05] first half of next year when we'll start [7:06] seeing some really robust evaluation [7:08] frameworks around around Agentic. Right [7:10] now most of the companies that are [7:12] working at least a peer group that I [7:13] have talked to they are kind of [7:15] augmenting their own knowledge with [7:17] external uh uh evaluation frameworks and [7:21] then collaboration typically you will [7:22] have a multi multi- aent scenario. So [7:24] how do agents collaborate? So how do you [7:26] manage the kind of let's say different [7:29] objectives of agents and and how do you [7:31] reconcile them? So that's another key [7:33] piece that that will have to be factored [7:35] in as you set up the foundation. [7:39] Now again this is a kind of a more of a [7:41] flow of of how you would set this up. [7:45] Typically you would have multiple [7:47] agents. Here I've considered two [7:48] different agent tech uh frameworks. Top [7:52] one that has agent one and two and the [7:53] bottom one that has three and four. So [7:55] these can be one can be maybe let's say [7:57] potentially from Azure. Second can be an [7:59] open source one. Now agents will have to [8:02] talk across these frameworks and Google [8:05] came up with this PL with this uh [8:07] protocol called A2A agent to agent which [8:09] allows this abstraction across different [8:12] agentic frameworks. It's it's it's good [8:15] but again the thing is one thing you'll [8:17] have to figure out is if you allow these [8:20] agentic frameworks to proliferate within [8:23] organization and you have to do this A2A [8:24] connections across multiple agents adds [8:26] to the overhead adds to the cost. So, so [8:29] it'll kind of you have to think about [8:31] simplicity and ease of use also. [8:34] Simplicity would be that you limit the [8:36] number of frameworks. Ease of use is [8:39] everybody can choose their their [8:41] framework easy to use but difficult to [8:43] manage and higher overhead. Then through [8:46] the language model they go through [8:48] basically they work with this framework [8:50] called uh model context protocol MCP. [8:54] MCP is nothing but an abstraction layer [8:57] between your agentic framework and the [9:00] resources that are needed to answer a [9:02] question. So when you're booking a [9:03] ticket to Italy, the previous example, [9:06] it has to it may go to Expedia. Now to [9:08] go to Expedia, it will go through MCP. [9:10] MCP MCP will route that call to Expedia. [9:13] It'll basically fill the forms on [9:15] Expedia for you. And if you give if you [9:18] want a fully autonomous agent, it can [9:20] even purchase a ticket for you within [9:22] your budget. [9:24] So really powerful but needs a lot of [9:27] careful definition and man and and [9:30] management. [9:32] So uh [9:35] now let's get a bit into MCP. So [9:39] fundamentally I think the question that [9:41] that you will do to you to basically ask [9:43] is that hey I have APIs I mean I use [9:48] let's say if I'm just using one one [9:50] external system why don't I just API [9:52] into it [9:53] very good question so the thing that you [9:56] have to figure out is how many external [9:58] resources how many data resources or [10:01] websites or or internal applications [10:03] does your agentic system need to access. [10:06] If there's a large number, it may be [10:09] better to create this abstraction layer [10:10] of MCP. If there's if there's very if [10:13] there are very few number of these [10:14] external resources that your system [10:16] needs to access, you may actually not [10:17] need an MCP. You can you should be you [10:19] you may be able to get uh get uh without [10:22] using an MCP. So that's something you [10:25] need to figure out. Now how an MCP works [10:28] is that basically it takes a basically a [10:31] question data agent throws to it through [10:34] a language model and basically it it [10:37] it's got two components. One's called [10:39] what's called MCP client. MCP client is [10:41] nothing but a list of all the resources [10:44] that the MCP has access to. So in this [10:48] case in the case of this uh uh this [10:51] Italy trip it will have access to that. [10:52] Okay, you're doing a travel booking. [10:54] potentially go to different travel [10:56] sites. It'll have a list of travel [10:57] sites. It'll have a list of maybe you [11:00] like to stay in Airbnb. So, it'll list [11:01] out all the Airbnbs and VBO options that [11:04] are available there. And it will also [11:07] look at kind of local tourism options in [11:09] terms of you want to take day trips [11:10] somewhere. So, it just keeps that list. [11:13] What MCP server does is once you've sent [11:17] a query that okay, I want to spend two [11:18] days in Rome. Find find me a place to [11:21] stay. Find what do I do? Then basically [11:24] it says okay 2 days in Rome so you need [11:25] a place to stay. Okay then it will send [11:28] the MCP server will send a send send a [11:30] request to VRB or Airbnb and try to [11:33] figure out which is at the right [11:35] location according to your your [11:38] requirements. So that's basically what [11:40] so that's the execution engine. MCP [11:43] client is uh basically just a kind of a [11:46] list of what resources are available to [11:48] be accessed. the server is on that [11:51] actually accesses those resources and [11:52] gets the answer and gives it back to the [11:54] agent. [11:55] So here as I said one key thing is [11:57] depending upon how many different [11:59] resources you want to access MCP versus [12:01] API then again how many MCP servers you [12:04] may want to set up configure different [12:06] MCP servers depending upon the different [12:08] there may be one for travel one for [12:10] finance one for let's say marketing and [12:12] so on and so forth depending upon how [12:14] you want to do it then token usage [12:17] extremely important by default MCP uh [12:21] ser MCP guzzles uh tokens so you will [12:24] have to one define your prompts very [12:26] carefully. Second, even the kind of [12:28] requests that MCP client the resource [12:32] listing you will as well as the request [12:34] that MCP server raises on the resources, [12:36] you will have to track them very [12:37] carefully because that's one of the [12:40] largest I would say consumption uh u [12:45] causes in in an agentic framework. So be [12:49] very careful about the token usage part [12:50] and and then again evaluation we talked [12:52] about and quality monitoring in terms of [12:54] the how the result is is coming out. So [12:57] that's uh on MCP [13:01] and then finally coming to language [13:04] model. So this often gets overlooked but [13:08] if you are looking at a three-year TCO [13:10] or a phops type of an approach the most [13:13] critical part this is where you spend [13:15] the most money in agentic framework [13:17] because every time anything is happening [13:20] there is a there's a token consumption [13:21] happening in terms of a request or in [13:24] terms of a response. So rather than [13:26] using commercial try to use a kind of a [13:30] distilled or fine-tuned uh llm language [13:34] model you don't necessarily need a large [13:36] language model even small language model [13:38] will do the trick. So these are some of [13:40] the techniques that that can be used uft [13:44] parameter efficient and uh smaller quant [13:46] quantization is something that's worked [13:47] very effectively. So uh yeah that's all [13:50] I know I ran through a lot but uh [13:52] believe me these are the questions you [13:54] should be asking your uh technology [13:56] teams and your providers uh uh service [13:59] providers. Thank you very much. [14:01] [Applause]