Adopt Automates API Tool Extraction to Replace Enterprise Web UIs

Architecting for AI Action: Building the Foundation for Autonomous Enterprise Agents With Adopt — Ai4

Ai4

By mapping user sessions directly to underlying network payloads, enterprises can bypass complex dashboard rebuilds and deploy autonomous agents that execute multi-step business actions under strict role-based access controls.

AI-assisted summary of Architecting for AI Action: Building the Foundation for Autonomous Enterprise Agents With Adopt — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 24, 2025, this presentation features Vijay Sagar, Founding AI Engineer at Adopt. According to the session description, the talk demonstrates how to architect enterprise infrastructure using composable API primitives that empower autonomous AI agents to perform secure, proactive actions across modern and legacy systems.

## TL;DR Vijay Sagar details the shift from static, form-heavy web user interfaces to proactive conversational agents that execute complex tasks on behalf of users. He outlines three non-negotiable enterprise guardrails—data accuracy grounded in APIs, strict contextual relevance, and role-based access control security—to safeguard brand trust. The talk also explores engineering and product hurdles in multi-agent orchestration and quality evaluation, concluding with an overview of how Adopt's platform autonomously discovers UI workflows, generates API tools from sandboxed user journeys, and provides continuous operational telemetry.

## Atmosphere & style The presentation carries a crisp, pragmatic, and builder-focused tone tailored for software architects and product leaders navigating production AI. Sagar speaks with steady, authoritative clarity, grounding high-level architectural concepts in tangible business scenarios like dynamic rental pricing and multi-tenant ad analytics. It feels like an informative and structured technical briefing designed to demystify how autonomous agents can reliably interface with legacy enterprise software.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Enterprise software is shifting from static, form-and-table user interfaces to proactive conversational agents that guide users and execute actions directly. - [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Conversational agents reduce cognitive overhead for infrequent users and unlock proprietary enterprise workflows without requiring extensive onboarding. - [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Shipping new capabilities through conversational endpoints allows organizations to deliver features faster without waiting for complete UI overhauls. - [04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) Production readiness for enterprise AI agents depends on three foundational guardrails: accuracy, relevance, and security. - [04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) Grounding agent recommendations directly in backend APIs and database state prevents hallucinations in critical operational decisions. - [04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) Agent execution privileges must strictly mirror authenticated user permissions and RBAC to maintain tenant isolation in multi-tenant platforms. - [08:08](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=488s) Scaling beyond single-purpose assistants requires multi-agent orchestration to route requests and manage contextual handoffs across domain-specific agents. - [08:08](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=488s) Automated evaluation pipelines, such as LLM-as-a-judge and regex benchmarking, are necessary to validate agent responses against ground-truth application data. - [08:08](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=488s) Product owners require fine-grained tone governance to adapt agent communication styles appropriately across distinct verticals like finance or healthcare. - [11:52](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=712s) Autonomous action discovery can explore web UIs, recording DOM states and network traffic to automatically generate executable API tools for agents. - [11:52](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=712s) Real-time telemetry and observability dashboards provide essential feedback loops to identify user friction and iteratively refine agent policies.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Transforming Complex Web UIs into Conversational Agents

Traditional web applications have long relied on static user interfaces—clicking buttons, filling out extensive forms, navigating dense tables, and applying filters. While this paradigm has driven software for decades, it inherently lacks intelligence. Vijay Sagar, Founding AI Engineer at Adopt, presents a vision to "build an agent for every web application," replacing tedious manual workflows with conversational interactions where software actively guides, advises, and performs tasks on behalf of the user.

```mermaid flowchart TD accTitle: Transition from Static UI to Conversational Agent subgraph Legacy["Traditional Web UI"] A["User Input: Forms & Filters"] --> B["Manual Navigation & Clicks"] B --> C["Static Application Output"] end subgraph Agentic["Adopt Conversational Interface"] D["User Natural Language Prompt"] --> E["Agent Intent & Domain Logic"] E --> F["Automated API Execution"] F --> G["Intelligent Advice & Proactive Action"] end ```

To illustrate the shift, Sagar shares real-world enterprise examples where domain complexity overwhelms end users: - **Short-Term Rental Pricing Platform:** In a B2B/B2C dynamic pricing application, property owners often struggle to calculate optimal nightly rates or seasonal zip-code occupancy benchmarks. Rather than forcing users through complex pricing tables, the agent diagnoses performance directly: noting that a property has 0% occupancy against an expected 33% benchmark, explaining that a listing was configured suboptimally, and recommending a price reduction from $100 down to $71 while asking, "Can I do that for you?" - **Market Research & Advertising Analytics:** A complex enterprise research platform typically requires hours of onboarding. When an employee receives a sudden Friday afternoon request from their manager to "generate a report on this market segment and have it on my desk by end of day," they no longer need to comb through documentation or master convoluted dashboard filters. They simply query the agent, which queries the data, performs the analysis, and delivers the finalized report.

Moving beyond static UIs liberates proprietary business IP, eliminates cognitive overhead for infrequent users who would otherwise face a recurring learning curve, and dramatically accelerates feature delivery cycles by launching new capabilities directly through conversational endpoints before investing in complete UI overhauls.

---

[04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) Foundational Guardrails: Accuracy, Relevance, and Security

Deploying user-facing enterprise agents places a company’s brand reputation and customer trust directly on the line. Ensuring production readiness requires building systems on three non-negotiable pillars:

```mermaid flowchart LR accTitle: Core Pillars of Enterprise Agents subgraph Pillars["Enterprise Agent Guardrails"] direction TB P1["Accuracy: Grounded in APIs & DBs"] P2["Relevance: Strict Task Alignment"] P3["Security: User-Level RBAC Mirroring"] end Pillars --> Prod["Trustworthy Production Agent"] ```

- **Accuracy:** Responses must be grounded in underlying databases, verifiable APIs, or explicit application logic. If an agent recommends setting a nightly rental rate to $71 or $120, that figure must reflect authoritative backend calculations rather than an LLM hallucination. - **Relevance:** The conversational scope must remain tightly bound to the user's immediate intent. If a customer is troubleshooting zero listing bookings, the agent must diagnose the issue rather than going off on unrelated promotional or upselling tangents. - **Security and Persona Grounding:** An agent's execution privileges must strictly mirror the identity and role-based permissions of the authenticated human user it represents. For example, if a multi-tenant ad management platform serves both Coca-Cola and Pepsi, an ad manager from Coca-Cola cannot ask the agent to "show me ad campaign performance for my competitor" and receive data. By anchoring agent tool calls to established enterprise access control lists (ACLs) and user session personas, multi-tenant isolation is maintained.

---

[08:08](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=488s) Scaling from PoC to Production Across Engineering and Product

Bridging the gap between an initial proof-of-concept (PoC) and a scaled enterprise deployment introduces distinct technical and operational challenges depending on stakeholder roles:

```mermaid flowchart TD accTitle: Multi-Agent Orchestration and Scaling UserQuery["User Request"] --> Orch["Agent Orchestration Layer"] subgraph Agents["Specialized Domain Agents"] Pricing["Pricing & Analysis Agent"] Compliance["Legal & Compliance Agent"] Maintenance["Maintenance & Upkeep Agent"] end Orch --> Pricing Orch --> Compliance Orch --> Maintenance Agents --> Eval["QA Benchmarking: LLM-as-a-Judge / Regex"] ```

#### Engineering Challenges (Scaling 1 to 100) - **Multi-Agent Orchestration:** Enterprise workloads quickly outgrow a single monolithic assistant. As teams add specialized agents (e.g., pricing analysis, legal and compliance, maintenance and upkeep), an orchestration layer must route incoming prompts, manage context, and seamlessly detect when a conversation boundary is reached to hand off context to another domain specialist. - **Quality Assurance & Automated Benchmarking:** Continuous testing pipelines must evaluate agent responses against ground-truth API outputs and database states, utilizing automated evaluation patterns such as LLM-as-a-judge or structural regex matching.

#### Product Owner Considerations - **Speed to Launch:** Because agentic software lacks legacy playbooks, product teams need rapid prototyping frameworks to get an initial agent into production quickly, validate strategic assumptions, and align cross-functional teams. - **Observability and Telemetry:** Real-time analytics dashboards are required to inspect user journeys, identify where dialogues fall short or cause user confusion, and iteratively tighten prompts and tool schemas. - **Tone and Personality Controls:** Agents represent the brand externally and require fine-grained stylistic governance tailored to the vertical—maintaining a concise, precise tone for financial workflows while adopting an empathetic, consultative posture in healthcare contexts.

---

[11:52](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=712s) The Adopt Integration Architecture and Workflow

The Adopt platform serves as an enterprise middleware layer that decouples application logic from agent runtime execution. The host company provides its proprietary backend—business logic, existing APIs, databases, custom agent pipelines, and third-party Model Context Protocol (MCP) servers. Adopt layers on action discovery, tool extraction, persona management, and multi-surface embedding options across web sidebars, homepages, or programmatic APIs.

```mermaid flowchart LR accTitle: Adopt Platform Integration Architecture subgraph Enterprise["Enterprise Backend"] KB["Knowledge Base"] APIs["REST / GraphQL APIs"] DB["Data Stores & MCPs"] end subgraph AdoptEngine["Adopt Platform Engine"] Crawler["Agent Exploration & Session Recording"] Extract["Action Discovery & Tool Extraction"] Dash["Observability & Analytics Feedback Loop"] end subgraph Surfaces["Deployment Endpoints"] UI["Web Sidebar / Homepage"] EmbedAPI["Programmatic APIs"] end

KB --> Crawler APIs --> Extract Crawler --> Extract Extract --> Surfaces Surfaces --> Dash Dash -.-> Extract ```

The onboarding and integration process operates through four key stages: 1. **Knowledge Ingestion:** The enterprise links its existing product documentation and knowledge base, allowing the platform to map core user journeys, key terminology, and application objectives. 2. **Autonomous Action Discovery:** Adopt is granted access to a sandboxed test account. An AI agent logs into the UI, executes the recorded user journeys, and captures DOM screenshots alongside network-level API requests. 3. **Automated Tool Generation:** By analyzing both the user-facing interface and the corresponding network payloads, the engine automatically extracts out-of-the-box API tools for every action supported by the web app, allowing builders to compose deep workflows that combine multi-API orchestrations with LLM interpretation. 4. **Telemetry and Feedback Loops:** Customer conversations are tracked via an analytics dashboard to evaluate completion rates, highlight edge-case failures, and iteratively refine agent policies.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Traditional web applications have long relied on static user interfaces—clicking buttons, filling out extensive forms, navigating dense tables, and applying filters. While this paradigm has driven software for decades, it inherently lacks intelligence. — Vijay Sagar

> [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) build an agent for every web application — Vijay Sagar

> [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Rather than forcing users through complex pricing tables, the agent diagnoses performance directly — Vijay Sagar

> [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Can I do that for you? — Vijay Sagar

> [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Moving beyond static UIs liberates proprietary business IP, eliminates cognitive overhead for infrequent users who would otherwise face a recurring learning curve — Vijay Sagar

> [04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) Deploying user-facing enterprise agents places a company’s brand reputation and customer trust directly on the line. — Vijay Sagar

> [04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) An agent's execution privileges must strictly mirror the identity and role-based permissions of the authenticated human user it represents. — Vijay Sagar

> [08:08](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=488s) Enterprise workloads quickly outgrow a single monolithic assistant. — Vijay Sagar

> [11:52](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=712s) The Adopt platform serves as an enterprise middleware layer that decouples application logic from agent runtime execution. — Vijay Sagar

> [11:52](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=712s) By analyzing both the user-facing interface and the corresponding network payloads, the engine automatically extracts out-of-the-box API tools for every action supported by the web app — Vijay Sagar

## Mentioned Vijay Sagar, Adopt, Ai4, Model Context Protocol (MCP), Coca-Cola, Pepsi, GraphQL, REST APIs

## Video description Vijay Sagar, Founding AI Engineer, Adopt

Abstracts In today’s fast-paced business environment, AI-powered agents have the potential to transform enterprise operations. However, to achieve this, we need more than just powerful models—we need a solid, reliable infrastructure. This talk will demonstrate how to architect systems with composable API primitives that enable AI agents to seamlessly interact with both legacy and modern APIs. These intelligent building blocks normalize data, ensure consistency, and allow AI to perform actions autonomously, securely, and reliably.

Key Takeaway: Learn how to design systems that don’t just deliver data to AI but empower agents to take intelligent, proactive actions across your enterprise, driving real business value.

[Source](https://www.youtube.com/watch?v=UI8oeLSa1ao)

中文简报

Adopt 自动化提取 API 工具以替代企业 Web UI 通过将用户会话直接映射到底层网络有效载荷，企业可以绕过复杂的仪表板重构，并部署在严格的基于角色的访问控制下执行多步骤业务操作的自主 Agent。 Vijay Sagar 详细阐述了从静态且表单繁琐的 Web 用户界面向代用户执行复杂任务的主动型对话式 Agent 的转变。他指出了维护品牌信任不可妥协的三大企业级护栏——基于 API 的数据准确性、严格的上下文相关性以及基于角色的访问控制安全性。演讲还探讨了多 Agent 编排和质量评估方面的工程与产品难题，最后概述了 Adopt 平台如何自主发现 UI 工作流、从沙箱用户旅程中生成 API 工具并提供持续的运营遥测。 企业级软件正从点击与填表的静态 UI 转向主动式对话 Agent，直接引导用户并执行业务操作，从而降低使用门槛并加速新功能发布。 (0:04) 生产级 Agent 依赖三大核心防线：基于后端 API 与数据库以确保准确性、紧密贴合任务意图以保证相关性，以及严格映射用户角色访问权限（RBAC）以维护多租户安全隔离。 (4:59) 从概念验证迈向规模化落地时，工程团队需依托多 Agent 编排机制实现跨领域任务路由与上下文移交，并借助 LLM-as-a-judge 等自动化评测手段持续验证输出质量。 (8:08) Adopt 架构通过沙盒探索自动化捕获 DOM 截图与网络 API 请求，将应用操作提炼为标准化工具，实现企业后端逻辑与 Agent 运行时的高效解耦集成。 (11:52)

Reading summary

Vijay Sagar, Founding AI Engineer at Adopt, argues that enterprise software must transition from static graphical user interfaces to conversational agents that autonomously execute backend actions. Successfully operationalizing these agents requires rigorous guardrails for data accuracy and security, alongside multi-agent orchestration and automated tool extraction. Replacing Complex Interfaces with Conversational Action (0:04) Traditional web applications rely heavily on manual navigation, multi-step forms, and static data tables, which Sagar contends creates high cognitive friction and steep learning curves for infrequent users. Conversational interfaces change this dynamic by shifting software from passive data displays into active advisors that diagnose domain-specific problems and execute remediation directly via backend APIs. To illustrate, Sagar highlights a dynamic short-term rental pricing tool where property owners often struggle to calculate optimal nightly rates. Instead of forcing users to analyze benchmark spreadsheets, an agent identifies an underperforming listing with 0% occupancy against a 33% local benchmark, advises dropping the rate from $100 to $71, and offers to make the adjustment automatically. A similar dynamic applies to analytics platforms, where natural language queries can assemble Friday afternoon research reports on demand. Sagar notes that this pattern accelerates feature delivery by exposing business capabilities conversationally before undertaking costly front-end redesigns. Ensuring Production Guardrails for Enterprise Deployment (4:59) Deploying customer-facing agents introduces brand and operational risks, which Sagar states must be mitigated through three core requirements: accuracy, relevance, and security. Accuracy mandates that recommendations are strictly grounded in authoritative backend databases, verifiable APIs, and deterministic application logic rather than unconstrained model predictions, preventing hallucinations in high-stakes actions like setting rental rates. Relevance keeps agent interactions focused strictly on resolving the user's immediate problem, preventing the model from wandering into unrelated upselling or extraneous topics. Security relies on enforcing role-based access control (RBAC) at the agent level by mirroring the authenticated user's session credentials. In multi-tenant systems—such as an ad platform serving competing brands like Coca-Cola and Pepsi—an agent bound to one tenant's access permissions cannot fulfill prompts requesting competitor analytics, preserving absolute data isolation. Operationalizing Multi-Agent Systems and Governance (8:08) Transitioning from proof-of-concept to production introduces distinct technical and operational challenges across engineering and product teams. Sagar observes that enterprise workloads quickly outgrow monolithic assistants, requiring an orchestration layer to route user requests across specialized domain agents—such as pricing, compliance, or maintenance—while managing state handoffs across conversational boundaries. Engineering teams must also implement continuous testing pipelines utilizing regex matching and LLM-as-a-judge patterns to evaluate outputs against ground-truth API data. Simultaneously, product managers must navigate the absence of established design playbooks for agentic software. This requires rapid prototyping frameworks to test strategic assumptions, telemetry dashboards to pinpoint conversational breakdowns, and fine-grained tone governance. For instance, product teams must ensure the agent communicates concisely in financial workflows while adopting a consultative, empathetic tone in healthcare settings. Automated Tool Extraction and Middleware Architecture (11:52) The Adopt platform functions as enterprise middleware that decouples application logic from agent runtime execution, integrating with host backends, REST or GraphQL APIs, databases, and Model Context Protocol (MCP) servers. Rather than requiring developers to manually build individual tool integrations, the platform automates action discovery by running an AI agent in a sandboxed environment to execute documented user journeys. By capturing front-end DOM states alongside network-level API requests, the system automatically extracts executable API tools for every supported web interaction. Builders can then compose complex, multi-API workflows executed across web sidebars, homepages, or programmatic APIs. Real-time telemetry dashboards monitor task completion rates and surface edge-case failures to continuously refine agent policies. Sagar concludes that building autonomous enterprise agents depends on automating the bridge between user-facing workflows and underlying APIs, but emphasizes that long-term production success requires strict adherence to user-level permissions and automated evaluation against verified enterprise data.

精炼阅读摘要

Adopt 初创 AI 工程师 Vijay Sagar 提出，企业级智能体正在重塑传统 Web 应用，通过将复杂表单与仪表盘交互转化为对话式自动化操作，在释放业务数据潜力的同时降低认知负担。然而，这一转变的成功落地高度依赖于以底层 API 为依托的准确性、严格的用户级权限隔离，以及跨专业智能体的编排与自动化工具生成流水线。 从静态界面转向对话式智能操作 (0:04) 传统的企业级 Web 应用长期依赖按钮点击、表单填写、密集数据表格及多级筛选等静态交互界面。Sagar 认为，这种模式固化了操作流程且缺乏原生智能，往往给偶发或初级用户带来沉重的使用门槛。以短租动态定价平台为例，房东常难以在复杂表格中测算最优每晚房费或区域入住率基准；智能体可直接诊断房源表现，明确指出在预期基准入住率为 33% 的情况下该房源入住率为 0%，并结合配置问题建议将价格从 100 美元调降至 71 美元，同时主动向用户确认是否代为执行调整。在市场调研与广告分析场景中，员工面临周五下午突发报告需求时，不再需要耗时查阅文档或设置复杂的控制台筛选器，而是直接向智能体提问并由其调用底层数据完成分析。 采用对话式界面还能使企业摆脱传统前端开发的周期制约。企业无需在验证初期耗费巨资对既有前端进行全面重构，而是可以直接通过对话端点对外交付新增的业务功能。这一转变不仅释放了企业私有的业务知识产权与数据价值，还有效消除了用户因系统不常用而反复产生的二次学习成本，大幅压缩了产品迭代交付周期。 生产级智能体的准确度与权限隔离护栏 (4:59) 将面向终端用户的企业级智能体推向生产环境，直接关系到品牌声誉与客户信任。Sagar 指出，生产环境必须牢固建立在三大核心支柱之上：第一是准确性，智能体的所有回答和业务建议必须严格锚定在底层数据库、可验证的 API 以及显式的业务逻辑之上。例如，智能体向租户推荐 71 美元或 120 美元的房费定价时，该数据必须源于权威的后端计算流程，杜绝由大语言模型（LLM）幻觉凭空捏造。 第二是任务相关性，对话范畴必须紧密受限于用户的即时意图。当客户正在排查房源零订单的故障时，智能体必须集中精力诊断核心问题，严禁偏离主题进行无意义的营销推广或交叉销售。第三是安全性与用户角色映射，智能体的执行权限必须严格继承并映射所代表的已认证用户的身份及基于角色的访问控制（RBAC）。以同时服务于可口可乐和百事可乐的多租户广告管理平台为例，可口可乐的广告经理绝不能通过智能体获取竞品广告活动的表现数据；通过将智能体的工具调用与既有企业访问控制列表（ACL）和会话身份深度绑定，系统得以在自动化操作中维持严密的多租户数据隔离。 多智能体编排扩展与工程产品治理 (8:08) 当系统从初始概念验证（PoC）阶段迈向企业级规模化部署时，工程研发与产品运营面临着截然不同的双重挑战。在工程技术层面，单一单体架构的智能体难以支撑复杂的企业级业务。随着定价分析、合规审查以及运营维护等专业领域智能体的引入，系统必须依赖中央编排层来分发用户意图、维护上下文状态，并在对话触及特定专业边界时无缝切换移交上下文。同时，质量保证（QA）流程需引入自动化基准测试机制，借助大模型作为裁判（LLM-as-a-judge）或结构化正则表达式，持续比对智能体输出与实际 API 返回值及数据库真实状态的一致性。 在产品管理与运营治理维度，由于智能体软件缺乏成熟的现成方案，产品团队亟需轻量级的快速原型开发框架以迅速上线并验证业务假设。为了保障运营安全，企业需要建立实时可观测性与遥测仪表盘，全流程审查用户交互轨迹，精准定位对话卡顿或产生歧义的具体环节，从而针对性微调 Prompt 和工具描述 Schema。此外，由于智能体代表品牌形象对外发声，产品端必须针对不同垂直行业设定差异化的话术基调与人格特质：在金融工作流中保持严谨精炼的语言风格，在医疗等敏感场景中则切换为富有同理心和咨询性质的交互姿态。 Adopt解耦架构与端到端工具生成 (11:52) Adopt 平台的定位是连接企业既有业务逻辑与智能体运行时的中间件层，其核心设计原则是实现应用逻辑与智能体执行的彻底解耦。企业客户仅需提供自身专有的后端基础设施，包括业务逻辑代码、REST 或 GraphQL API、数据库、自定义智能体流水线以及第三方模型上下文协议（MCP）服务器；Adopt 则负责在之上叠加操作发现、工具提取、角色管理以及多终端嵌入能力，支持将智能体无缝嵌入到 Web 侧边栏、系统主页或以编程方式通过 API 暴露。 系统上线与集成的全流程分为四个规范化阶段：首先是知识导入阶段，企业接入现有的产品说明文档和知识库，供平台梳理核心用户旅程、关键术语与业务目标；其次是自主操作发现阶段，AI 智能体获得沙盒测试账户授权后登录系统前端界面，自动运行预设的用户旅程，同步捕获 DOM 树截图与网络层 API 请求；随后进入自动化工具生成阶段，引擎交叉分析前端 UI 变化与底层网络数据包，自动为 Web 应用所支持的每项操作提取开箱即用的 API 工具，使开发者能够快速组合多 API 调度与大模型推理的高阶工作流；最后依托遥测与反馈闭环阶段的分析仪表盘，持续监测任务完成率并暴露边界异常，驱动智能体执行策略的迭代优化。 Vijay Sagar 描绘了通过中间件将 UI 操作与底层 API 自动提取为工具、进而构建自主企业级智能体的技术路径，但该模式的有效性仍高度依赖企业既有知识库与 API 的完备性，如何在高度定制化且缺乏规范接口的多租户企业环境中保障多智能体协作的绝对可靠，仍是该架构在实际推广中面临的持续挑战。

Detailed notes — zh

## Background 该内容由 Ai4 频道于 2025 年 8 月 24 日发布在 YouTube 平台，主讲人为 Adopt 创始 AI 工程师 Vijay Sagar。视频简介指出，演讲聚焦于如何通过可组合的 API 原语构建坚实可靠的基础设施，使 AI 智能体能够安全、自主地连接企业新老系统并主动执行操作，从而驱动实际业务价值。 ## TL;DR 演讲系统阐述了将传统 Web 应用的“点击与填表”模式演进为双向自主 Agent 交互的架构路径，并结合短租定价与市场分析等场景展示了对话式 Agent 消除学习曲线、加速价值交付的优势。主讲人深入剖析了面向用户的生产级智能体必须坚守的准确性、相关性及角色权限隔离三大支柱，同时梳理了多 Agent 编排、基准测试与人设控制等规模化挑战。最后，演讲介绍了 Adopt 平台的端到端赋能方案，通过接入知识库与自动化模拟用户旅程捕获 API，帮助企业开箱即用生成 Agent 能力并实现监控闭环。 ## Atmosphere & style 整体氛围务实专业且极具前瞻性，呈现出高浓度技术分享与商业路演相结合的快节奏风格。讲者表达清晰流畅、直奔业务痛点，通过生动的实际案例将复杂的工程编排与安全隔离逻辑拆解得通俗易懂。整场演讲洋溢着对 Agentic 交互范式变革的笃定与热情，带给听众清晰的落地路径与强烈的实战信心。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) 传统 Web 应用程序长期依赖点击、填表和筛选等机械交互，Adopt AI 的目标是通过为每个应用构建智能体，将其升级为具备双向对话、主动诊断与任务执行能力的智能交互界面。 - [01:13](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=73s) 在短租定价和市场分析等复杂场景中，智能体不仅能提供最优定价等主动决策建议，还能直接通过对话接口交付分析报告，显著消除低频使用软件的学习成本。 - [03:10](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=190s) 通过对话接口发布功能可以让企业绕过漫长的 UI 开发排期，直接交付核心业务知识产权（Business IP）并快速验证用户价值。 - [04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) 构建面向终端用户的生产级智能体必须坚守三大支柱：输出必须严格锚定真实数据的准确性、对话不跑偏的相关性，以及精准映射登录角色的安全性。 - [07:00](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=420s) 在企业级多租户系统中，智能体的权限边界必须与用户角色实现 1:1 精准隔离，杜绝大模型跨租户越权访问竞对数据。 - [08:08](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=488s) 当企业智能体从 1 个规模化扩展至 100 个时，工程团队需攻克多 Agent 意图路由与平滑上下文移交，并引入 LLM-as-a-judge 等基准测试保障准确率。 - [09:40](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=580s) 产品负责人需要兼顾冷启动上线速度、全链路对话可观测性，以及针对不同行业（如金融的干脆直奔主题、医疗的同理心）进行精细的语气与人设控制。 - [11:45](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=705s) Adopt 平台通过解耦架构，让企业自带 API、业务逻辑与 MCP 工具，平台侧负责自动化的工具发现、人设管理以及侧边栏/首页/API 等全渠道部署。 - [12:32](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=752s) 企业只需接入知识库并提供测试账号，Adopt 智能体便能自主模拟用户旅程、捕获操作截图与 API 调用，自动掌握并生成应用的全套操作技能。 - [13:15](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=795s) Adopt 提供了多 API 串联编排能力与监控分析看板，使企业能够可视化追踪会话质量并快速收紧安全边界，形成持续优化的数据反馈闭环。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) 重新定义 Web 交互：为每一个应用程序注入智能体 长期以来，人们使用 Web 应用程序的标准方式始终停留在“点击按钮、填写表单、浏览数据表格以及配置各类筛选器”上。这种交互范式虽然沿用了数十年，却始终缺失了一个核心要素——**智能化（Intelligence）**。Adopt AI 的使命正是“为每一个 Web 应用程序构建专属智能体（Build an agent for every web application）”。当传统的图形界面转变为能够双向交流的智能对话界面时，用户不再需要机械地摸索功能，而是可以直接向系统提问、下达任务指令，甚至让应用程序主动提供操作指导、诊断建议与决策辅助。 ```mermaid flowchart LR accTitle: 从传统Web交互到智能体对话交互的演进 subgraph Traditional["传统 Web 交互"] A1["点击按钮 / 填表"] --> A2["复杂筛选与菜单"] --> A3["被动数据呈现"] end subgraph Agentic["智能体交互 (Adopt AI)"] B1["自然语言提问 / 指令"] --> B2["双向推理与主动诊断"] --> B3["自主执行与价值交付"] end ``` --- ### [01:13](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=73s) 落地案例：短租智能定价与复杂市场调研报告 在实际企业服务中，这种智能体范式已经展现出颠覆传统 UI 的价值： * **短租房源定价引擎（B2B / B2C 模式）**：房东的核心诉求是确定最优房价以及掌握特定邮政编码（Zip Code）区域的季节性入住率。传统的定价逻辑与界面极其繁琐。通过接入智能体，用户可以直接提出诉求：“帮我为这处房源设定合适的价格”。智能体会基于底层数据进行诊断并回复：“您当前设置的每晚价格是 100 美元，但系统测算的最优价格应为 71 美元。如果调整为 71 美元，将带来更好的收益表现，需要我帮您直接修改吗？”当用户遇到困境抱怨“完全没有预订”时，智能体能迅速定位原因：“您的实际入住率为 0%，而该区域同期的基准入住率应为 33%。问题出在房源配置存在次优项，如果关闭某项特定设置，即可改善转化。” * **广告与市场调研分析平台**：面对复杂的行业数据和陡峭的软件学习曲线，如果员工在周五下午突然接到主管要求“在下班前生成某细分市场表现报告”的紧急任务，传统做法需要耗费大量时间阅读知识库（Knowledge Base）、登录系统反复摸索。而借助面向用户的智能体，用户只需输入自然语言分析指令，系统便会自动调取数据、完成深度分析并直接交付成果。 将企业的核心业务知识产权（Business IP）和垂直领域专长从复杂的 UI 菜单中解放出来，赋予其对话能力，具有显著的业务优势： - **消除学习曲线**：避免用户因低频使用（每周或每月一次）而产生认知过载和反复遗忘。 - **加快价值交付周期**：开发出新功能后无需等待完整的 UI 开发上线流程，可通过对话接口先行发布、交付价值，并根据用户反馈再迭代界面。 --- ### [04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) 面向用户智能体的三大核心支柱：准确性、相关性与严格安全边界 构建面向终端用户（User-facing）的智能体与内部实验有着本质区别，因为“公司的品牌声誉直接系于其上，风险与赌注极其高昂”。要构建一个可信赖的生产级智能体，必须牢牢把握三个核心维度： ```mermaid flowchart TD accTitle: 面向用户智能体的三大核心支柱 subgraph Pillars["企业智能体核心支柱"] Acc["准确性 (Accuracy)<br/>锚定真实数据 / 数据库 / API"] Rel["相关性 (Relevance)<br/>聚焦用户当前诉求，不跑偏"] Sec["安全性 (Security)<br/>权限精确映射至登录用户角色"] end Acc --- Rel --- Sec ``` 1. **准确性（Accuracy）**：智能体的所有回答必须与现实严格对齐，具有明确的事实依据（Grounded）。如果系统向房东建议定价 120 美元，该数值必须能在底层数据库或对应 API 中得到校验，绝不能由底层大语言模型（LLM）凭空臆造。 2. **相关性（Relevance）**：对话必须紧扣主题。当用户在排查房源没有预订的问题时，智能体不能脱离上下文强行推销其他增值服务。 3. **安全性与权限对齐（Security & Persona Isolation）**： - 智能体的操作权限与数据访问范围，必须与**当前登录的具体人类用户角色（Persona / Role）**实现 1:1 的精准对齐。 - 以多租户广告投放管理系统为例，假设平台同时服务了可口可乐（Coca-Cola）与百事可乐（Pepsi）。如果可口可乐的广告经理向分析智能体发出指令：“展示我竞争对手的广告投放表现”，系统必须通过标准化的权限体系直接拦截，杜绝 LLM 越权提取跨租户数据。 --- ### [08:08](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=488s) 规模化演进：工程编排、评测基准与产品化把控 从概念验证（PoC）走向支撑 1 到 100 个智能体的生产级系统，工程师与产品负责人（Product Owner）各自面临关键挑战： | 关注维度 | 角色视角 | 核心挑战与架构策略 | |:--- |:--- |:--- | | **多智能体编排 (Orchestration)** | 工程师团队 | 随着定价分析、合规审计、运维维护等多个垂直 Agent 的增加，系统需要精细的路由编排层，以准确识别用户意图并在各 Agent 之间平滑移交（Handoff）上下文。 | | **质量保证与基准测试 (QA & Benchmarks)** | 工程师团队 | 将智能体输出与数据库真实数据、标准 API 结果进行持续比对，采用 LLM-as-a-judge（大模型作为裁判）或正则表达式评测机制以保障准确率。 | | **快速上线速度 (Speed to Launch)** | 产品负责人 | 在缺乏通用标准规范的探索期，需以最快速度推出首个 Agent，从而验证业务策略并凝聚团队方向。 | | **可观测性与分析 (Observability & Analytics)** | 产品负责人 | 监控所有对话路径，识别高价值交互链路，找出对用户造成困惑或未产生价值的会话并持续调优。 | | **语气与人设控制 (Tone & Personality)** | 产品负责人 | 智能体直接代表企业对外形象。例如金融类 Agent 必须精准、干脆、直奔主题，需具备精细的 Persona 控制能力。 | ### [11:45](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=705s) Adopt 平台架构与端到端赋能 在实际业务场景中，不同领域的 Agent 需要具备不同的行为偏好与风格，例如医疗健康领域的 Agent 就必须表现得更加富有同理心与助人倾向，这就要求系统必须提供细粒度的控制能力。Adopt 平台正是为了帮助客户实现这一目标而构建的。在整个架构中，客户的既有应用位于左侧，企业只需提供自己的产品业务逻辑、底层 API、数据存储、已有的自研 Agent 以及第三方的 MCP（Model Context Protocol）工具集。 Adopt 平台则负责处理中间所有的复杂脏活：包括自动化动作发现、工具提取、设定 Agent 人设（Agent Personas），以及灵活的部署方式。企业可以自由选择将 Agent 嵌入到侧边栏（Sidebar）、主页（Homepage）或直接以 API 的形式对外提供服务，最终让终端用户能够无缝与智能体进行交互。 ```mermaid flowchart LR accTitle: Adopt 平台端到端架构流程图 subgraph Client["企业客户侧 (Bring Your Own)"] A["业务逻辑 & API"] B["数据存储 & 自研 Agent"] C["第三方 MCPs"] end subgraph Adopt["Adopt 平台核心能力"] D["动作与工具自动发现"] E["Agent 人设与行为控制"] F["多端部署能力\n(侧边栏 / 首页 / API)"] end subgraph User["终端用户交互"] G["终端用户 (End User)"] end Client --> Adopt Adopt --> User ``` - **客户自带资产**：企业自身的业务逻辑、API 接口、数据库、自研 Agent 实验成果及第三方 MCP 集成。 - **Adopt 核心赋能**：自动化工具抽取、人设与控制策略配置、全渠道部署通道。 --- ### [12:32](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=752s) 自动化学习流程与持续优化闭环 将应用接入 Adopt 的设置过程极其简单高效。首先，企业只需将 Adopt 接入现有的知识库，系统便会自动解析企业所在的业务领域以及关键的用户旅程（User Journeys）。随后，企业只需提供一个测试环境账号的登录凭证，Adopt 的 AI Agent 就会自主登录系统，端到端地模拟并重现这些用户旅程。在这个过程中，Agent 会自动收集操作截图并抓取背后的 API 调用，从而彻底掌握两件核心要素：“你的应用程序能做什么，以及它是如何做到的（what your application does and how it does it）”。 基于这些自动学习到的知识，Adopt 能够开箱即用地为企业生成初始的 Agent 技能集。这意味着用户在前端界面上能够执行的任何操作，都可以直接转化为 Agent 的能力。在此基础上，开发者还能进一步构建深度的 Agent 体验：组合调用多个 API，利用 LLM 对各 API 的返回结果进行综合推理与重组，并向用户呈现最终处理结果。与此同时，Adopt 提供了专门的监控看板（Dashboarding），企业可以清晰查看用户与 Agent 的真实对话记录、追踪成功交付的业务结果，并发现需要对 Agent 进行“收紧和规范”的薄弱环节，从而形成持续迭代与改进的数据反馈闭环。 - **快速接入四步法**：接入知识库解析业务上下文 $\rightarrow$ 提供测试账号 $\rightarrow$ Agent 自动复现旅程并捕获 API 与截图 $\rightarrow$ 开箱即用生成 Agent 能力。 - **深度智能化拓展**：支持多 API 串联编排与 LLM 智能结果解析。 - **监控与反馈闭环**：通过看板洞察对话质量与转化成果，持续微调和优化 Agent 行为边界。 --- ### [13:50](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=830s) 展会互动与案例体验 在演讲尾声，Vijay Sagar 热情邀请现场观众前往展台进行更深入的交流。Adopt 在会场设有 227 号展位，现场准备了精美礼品（Swag）并安排了相关互动活动。屏幕上也提供了二维码，参会者可扫码查看 Adopt 帮助其他客户实现业务落地的深度案例研究（Case Study），或直接预约专属的产品演示（Demo）。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=4s) Build an agent for every web application — Vijay Sagar > [01:13](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=73s) 帮我为这处房源设定合适的价格 — Vijay Sagar > [01:13](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=73s) 您当前设置的每晚价格是 100 美元，但系统测算的最优价格应为 71 美元。如果调整为 71 美元，将带来更好的收益表现，需要我帮您直接修改吗？ — Vijay Sagar > [01:13](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=73s) 您的实际入住率为 0%，而该区域同期的基准入住率应为 33%。问题出在房源配置存在次优项，如果关闭某项特定设置，即可改善转化。 — Vijay Sagar > [01:13](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=73s) 在下班前生成某细分市场表现报告 — Vijay Sagar > [04:59](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=299s) 展示我竞争对手的广告投放表现 — Vijay Sagar > [12:32](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=752s) what your application does and how it does it — Vijay Sagar > [12:32](https://www.youtube.com/watch?v=UI8oeLSa1ao&t=752s) 收紧和规范 — Vijay Sagar ## Mentioned Vijay Sagar, Adopt, Adopt AI, Ai4, Model Context Protocol (MCP), Coca-Cola, Pepsi ## Video description Vijay Sagar, Founding AI Engineer, Adopt Abstracts In today’s fast-paced business environment, AI-powered agents have the potential to transform enterprise operations. However, to achieve this, we need more than just powerful models—we need a solid, reliable infrastructure. This talk will demonstrate how to architect systems with composable API primitives that enable AI agents to seamlessly interact with both legacy and modern APIs. These intelligent building blocks normalize data, ensure consistency, and allow AI to perform actions autonomously, securely, and reliably. Key Takeaway: Learn how to design systems that don’t just deliver data to AI but empower agents to take intelligent, proactive actions across your enterprise, driving real business value. [Source](https://www.youtube.com/watch?v=UI8oeLSa1ao)

Source quotes

[0:04] Vijay Sagar: Build an agent for every web application [1:13] Vijay Sagar: 帮我为这处房源设定合适的价格 [1:13] Vijay Sagar: 您当前设置的每晚价格是 100 美元，但系统测算的最优价格应为 71 美元。如果调整为 71 美元，将带来更好的收益表现，需要我帮您直接修改吗？ [1:13] Vijay Sagar: 您的实际入住率为 0%，而该区域同期的基准入住率应为 33%。问题出在房源配置存在次优项，如果关闭某项特定设置，即可改善转化。 [1:13] Vijay Sagar: 在下班前生成某细分市场表现报告 [4:59] Vijay Sagar: 展示我竞争对手的广告投放表现 [12:32] Vijay Sagar: what your application does and how it does it [12:32] Vijay Sagar: 收紧和规范

Source quotes

[0:04] Vijay Sagar: Traditional web applications have long relied on static user interfaces—clicking buttons, filling out extensive forms, navigating dense tables, and applying filters. While this paradigm has driven software for decades, it inherently lacks intelligence. [0:04] Vijay Sagar: build an agent for every web application [0:04] Vijay Sagar: Rather than forcing users through complex pricing tables, the agent diagnoses performance directly [0:04] Vijay Sagar: Can I do that for you? [0:04] Vijay Sagar: Moving beyond static UIs liberates proprietary business IP, eliminates cognitive overhead for infrequent users who would otherwise face a recurring learning curve [4:59] Vijay Sagar: Deploying user-facing enterprise agents places a company’s brand reputation and customer trust directly on the line. [4:59] Vijay Sagar: An agent's execution privileges must strictly mirror the identity and role-based permissions of the authenticated human user it represents. [8:08] Vijay Sagar: Enterprise workloads quickly outgrow a single monolithic assistant. [11:52] Vijay Sagar: The Adopt platform serves as an enterprise middleware layer that decouples application logic from agent runtime execution. [11:52] Vijay Sagar: By analyzing both the user-facing interface and the corresponding network payloads, the engine automatically extracts out-of-the-box API tools for every action supported by the web app

Original transcript

[0:04] Welcome to this talk building the [0:06] foundation for autonomous enterprise [0:08] agents. [0:09] My name is Vijay. I'm from adopt AI. [0:13] Our mission is to build an agent for [0:16] every web application. [0:19] Now that begs the question, what kind of [0:21] agent am I talking about and why is it [0:24] important? [0:26] So consider any web application that you [0:28] have used recently right presumably you [0:31] got things done by clicking buttons [0:34] filling out forms [0:36] uh tables selecting filters right [0:42] now this has been the way we have built [0:45] web applications for many years but it [0:47] lacks one important thing intelligence. [0:51] What if you could talk to the web [0:52] application [0:54] right? What if you could ask questions? [0:56] What if you could ask it to do things [0:59] for you? [1:01] What if your application could talk to [1:02] you, maybe guide you, maybe advise you, [1:06] assist you? What if we could do all of [1:08] that? What would that experience look [1:10] like? [1:13] And so, here's a customer who we are [1:15] helping with their agentic journey. They [1:18] are a pricing engine for short-term [1:20] rentals, right? They are a B2B B2C model [1:24] and their users are people who own bunch [1:29] of rental properties and they're trying [1:31] to answer the question, what's the best [1:33] price for my property? What's the [1:35] typical occupancy rate at this time of [1:37] the year in this zip code, right? And [1:41] pricing is complex and their UI reflects [1:44] that complexity, right? But what if I [1:47] you are a user of this uh application [1:52] and you could just talk to them. You [1:54] could just bring your problem. Help me [1:56] set the rise right price for this [1:59] property. What is the right price? [2:02] Right? And the application talks to you [2:04] and it says well you have set it to $100 [2:07] but we believe the right nightly price [2:10] is $71. Right? And if you set it to $71, [2:14] these are the positive outcomes you [2:15] would reach. Can I do that for you? [2:17] Right [2:19] there. There's a different conversation. [2:22] A presumably unhappy user says, "Well, [2:25] I'm not getting any bookings. What do I [2:27] do?" And then the applications help you [2:30] helps you out. It tries to diagnose the [2:32] problem. It says, "Well, you are right. [2:34] Your occupancy has been at 0%." [2:37] When it should be 33%. Well, here's the [2:41] problem. you have set up the listing uh [2:44] suboptimally. Maybe if you turn off this [2:46] setting, you would get better outcomes, [2:49] right? And that is the application [2:51] talking with you, helping you on your [2:53] journey, helping the user on their [2:55] journey, right? Here's a different uh [2:59] customer again. We help them build their [3:02] agentic experience and [3:05] uh you know, they are in the advertising [3:08] uh and market research space. complex [3:11] app because the domain is complex, [3:13] right? Imagine you are a user of their [3:15] platform. Your company purchased the [3:16] subscription. [3:18] You haven't had the time to get used to [3:21] the complexity and your boss comes in [3:23] Friday afternoon and says, "Generate a [3:26] report on this market segment and have [3:28] it on my desk by end of day." Right? [3:31] What do you do? go to their knowledge [3:33] base, read it up, log into the app, uh [3:37] try to figure out how it works or just [3:40] talk to it, ask it the question that uh [3:44] you have in mind. Analyze the [3:45] performance of something. It pulls the [3:48] data, performs an analysis and gives you [3:51] an outcome. Job done. Right? That's sort [3:54] of why userf facing agents are [3:57] important. That's why we are solving [3:58] this problem. [4:00] You have unique domain expertise, [4:03] you have business IP. Why should you [4:06] confine it to a UI? Why not set it free [4:09] form of a text uh interface where users [4:12] can talk to it, have a conversation with [4:14] it, right? And it has a lot of [4:17] advantages, right? Complex UIs users [4:20] have to learn. It's cognitive overload [4:22] for them. An application that they use [4:24] once a month or once a week, they tend [4:27] to forget. and then they have to relearn [4:29] every time. So you skip the learning [4:31] curve. Your application is an expert. It [4:34] helps them on their journey. It talks to [4:35] them, walks them through the problem. [4:38] And then faster value delivery. You [4:40] build out new functionality. You don't [4:42] have to wait for the full uh UI cycle. [4:45] You can launch it via your [4:47] conversational interface, make it [4:49] available, deliver value, and based on [4:51] feedback, you build out the full UI [4:53] cycle. So that's the promise of agents [4:57] right [4:59] which brings us to the meat of the talk [5:02] what are the challenges in agent [5:04] building right remember we are building [5:06] userf facing agents right [5:10] company's reputation and brand value is [5:13] on the line and it the stakes don't get [5:15] any higher than this and so what are the [5:17] ingredients or what are the important [5:19] things that you need to keep in mind as [5:21] you embark on your own you know agent [5:24] building journey. By the way, how many [5:25] of you folks are, you know, considering [5:28] building agents or have already started [5:30] down this path? Show of hands, please. [5:33] Cool. Awesome. Well, uh I'm sure we [5:37] would love to connect and uh maybe for [5:39] the folks who aren't on that journey, I [5:41] hope to plant some thoughts in your [5:43] mind. Right? So, these are essential for [5:46] a successful agent, right? Accuracy, [5:49] responses must align with reality. They [5:51] must be grounded in something. So if you [5:54] were to ask your agent, what's the right [5:55] price for my rental property? And it [5:57] says $120, it must be backed by [6:00] something. There must be a UI or a [6:03] database where you can go check that is [6:06] it indeed 124 $120 or is it something [6:10] that you know an LLM backing your agent [6:14] made up based on what it imagines, [6:17] right? So responses must align with [6:20] reality. [6:22] Relevance, you should stay on topic. If [6:24] a user complains about not getting any [6:26] listings, you cannot go off on a tangent [6:28] and try to upsell something, right? And [6:33] the most important is security, right? [6:36] Uh here I'm going to make a [6:37] recommendation that whatever agent you [6:40] build its privileges must exactly match [6:45] a human user who it claims to represent. [6:49] Right? For example, if you are trying to [6:51] build now let me make up an example. [6:55] Imagine you're building an ad campaign [6:56] management application and you signed up [6:59] Pepsi and Coca-Cola. Congratulations. [7:02] Now you're building an agent, an [7:05] analysis agent, right? What happens if [7:08] someone from Coca-Cola comes in? By the [7:11] way, we are not affiliated with either [7:12] of these two, just picking them from for [7:14] recall value. What if someone from [7:16] Coca-Cola comes in and says, "Show me ad [7:21] campaign performance for my competitor." [7:24] How do you prevent the LLM from fetching [7:26] the data, exposing it to your [7:28] competitor? Right? And so you got to tie [7:30] it to some persona. And what is that [7:33] persona? That's a subtle point. You have [7:36] to tie it to [7:39] the persona that's logged in, right? [7:42] That person from Coca-Cola that uh let's [7:46] let's call it a role. An ad manager from [7:47] Coca-Cola is the person on whose behalf [7:50] you're accessing data. And if you ground [7:53] that, you follow standard security rules [7:55] that we have been following for many [7:57] many years while building web [7:58] applications [8:00] and you give the same access permissions [8:02] and the privileges to an agent, you'll [8:04] be fine for the most part. [8:08] Uh scaling up well now you have built [8:11] out a P, right? But there is a huge huge [8:15] value between going from a P to a [8:17] production agent, right? Now, if you are [8:19] an agent builder and you are an engineer [8:22] persona, right? How many of you folks [8:24] are here? Quick show of hands, please. [8:28] Oh, not too many. Well, but if you are, [8:31] so these are the things that would be [8:33] top of mind. [8:35] Going from one to 100, right? Um, we [8:38] have built out an agent, say a pricing [8:41] analysis agent. Now, we are not going to [8:44] stop there. We are going to build out a [8:46] legal and compliance agent. You're going [8:49] to build out a maintenance and upkeep [8:51] agent. You're going to build out a bunch [8:54] of other agents. And this is not you're [8:58] you're not going to stop. You're going [8:59] to keep on adding value by building more [9:01] and more agents. Now, how do these work [9:03] together? You're going to have to build [9:05] out an agent orchestration layer which [9:08] ensures that a user query goes to the [9:10] right agent. At some point, the [9:12] conversation is going to end, right? and [9:15] it's going to go to say legal and [9:17] compliance agent territory. So you need [9:20] to detect when a conversation ends with [9:22] a pricing agent and when it redirects to [9:25] a compliance agent, right? So you need [9:28] to build out that layer, right? Um then [9:32] there is quality assurance of course [9:33] benchmarks going back to accuracy. If a [9:37] user says what is the right price for my [9:38] property, then your benchmark should [9:41] compare your agent's response with what [9:43] it can pull from a database or an API. [9:46] It needs to compare alternates to ensure [9:48] that the responses match. And you could [9:52] probably use LLM as a judge. Even a [9:54] simple reax would do. But you do need to [9:57] have this. [10:00] Then if you are a builder with product [10:02] owner in your title, right? Do we have [10:05] those folks here? Show of hands again. [10:08] Okay, you. So you would be worried about [10:12] the speed of launch. Like we get it. We [10:15] get it that agentic building agents is a [10:18] long-term commitment. There is no [10:20] playbook for this. We are all doing this [10:22] for the first time. We know how high the [10:24] stakes are and so we do expect that [10:26] there will be a lot of discussion around [10:28] the right strategy [10:30] um the right tactics, how to get the [10:32] right team in place, right? And so speed [10:35] to launch, how do you get your first [10:36] agent really quickly out there so that [10:38] you can uh crystallize your strategy, [10:41] you can get your team to rally behind [10:43] this is one of your main concerns. [10:46] Then let's get let's say you get your [10:48] first agent out there. Then you need to [10:50] think of these things observability [10:52] and analytics, [10:55] right? How do we find out which of your [10:57] user journeys uh which of your agentic [11:00] conversations are most relevant and [11:02] useful to the user? [11:04] How do you find out which of those [11:06] conversations aren't adding value or [11:08] confusing the user? How do you tighten [11:10] up your agent responses? [11:14] Right? And tone and personality. And [11:16] this is most important from a product uh [11:18] owner point of view. Right? Your agent [11:20] must represent your company, your [11:23] company's values, the domain in which [11:26] your company works in and [11:31] uh it it's representing you to the [11:33] outside world. So you need fine controls [11:36] on the tone and persona, right? A [11:39] finance agent, it should be direct to [11:40] the point. a uh [11:45] healthcare agent, it needs to be more [11:47] empathetic and more helpful, right? So, [11:49] you need fine grain controls over that. [11:52] This is the adopt platform. This is uh [11:54] how we help our customers get to [11:56] outcomes. So, your application sits on [11:59] the left hand side. You bring the [12:01] product business logic. You bring your [12:02] APIs. You bring your data stores, your [12:05] own agents and if you have already u you [12:08] know started on your own agenting [12:09] journey uh your own agents and your [12:12] third party MCPs right everything else [12:15] adopt brings discovery of actions tools [12:19] extraction agent personas where and how [12:23] you choose to deploy your agent you know [12:25] sidebar homepage APIs [12:28] and then the end user gets to interact [12:30] with the agent [12:32] there a quick demo on how you set it up, [12:35] how it all comes together. So you point [12:39] us to your knowledge base which is where [12:41] we figure out what you do, which [12:42] business you are in, what your critical [12:44] user journeys are. Once you have done [12:47] that, you give us credentials to a test [12:49] account. [12:50] Our AI agent logs in, tries to create [12:53] those, recreates those user journeys. It [12:56] gathers screenshots and API calls. And [12:59] now it knows two things. what your [13:01] application does and how it does it. [13:04] With this knowledge, we can surface an [13:06] initial set of capabilities for you. Um [13:10] the API calls are all agentic [13:12] capabilities that come out of the box. [13:14] Anything that you do on the app can be [13:17] done via the agent and you can go from [13:19] there. You can build further. You can [13:22] create your own agents deep agentic [13:24] experiences which combine multiple APIs [13:27] you know have an LLM interpret the [13:29] results of the API combine them in [13:31] different ways present the outcomes to [13:33] the user then this is the dashboarding [13:36] experience where you can uh check on the [13:38] conversations that your users are having [13:41] successful outcomes where you need to [13:43] tighten up the agent right and that [13:45] becomes a part of your feedback loop [13:47] that you can use to improve upon agent [13:49] continuously [13:50] Right, that brings us to the end of the [13:53] talk. We have a booth at 2:27. [13:57] Uh, let me there's a there should be a [14:00] QR code here. We have some cool swag and [14:03] events. So, do come see us. Uh, let me [14:09] Yeah, there's a QR code. Uh, and there's [14:13] a link to a case study where you can [14:15] take a look at how we helped our other [14:17] customers and you can also schedule a [14:19] demo. [14:21] Right.