Kagen Deploys Governed AI Fleets to Replace Billed IT Service Hours

AI and Managed Services: Future Models of Symbiosis with AI — Ai4

Ai4

Managed service providers will fail unless they transition from selling human labor to operating governed agent fleets with runtime-enforced guardrails that slash execution costs.

AI-assisted summary of AI and Managed Services: Future Models of Symbiosis with AI — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel in August 2026, this session features Kanav Kahol, Chief AI Officer at Kagen.ai. According to the description, the presentation explores how artificial intelligence is shifting managed services from a labor-driven model to an intelligence-driven one, focusing on AI governance, token cost optimization, and sustainable frameworks for human-agent symbiosis.

## TL;DR Kanav Kahol presents an architectural blueprint for enterprise AI symbiosis, arguing that the future of managed services lies in operating fleets of governed agents rather than billing for human hours. He analyzes five core failure modes of early enterprise AI—amnesia, lack of rollback, hallucinations, opacity, and ungoverned execution—and demonstrates how they are resolved through runtime-enforced operating system guardrails. The session details the full-stack architecture of the Kira execution harness, the Kagen governance engine, and on-premises 'AI in a Box' appliances, highlighting how dynamic model routing and multi-agent swarm debates slash operational costs while ensuring deterministic human-in-the-loop oversight.

## Atmosphere & style The presentation carries an authoritative, pragmatic, and systems-engineering tone that directly cuts through generative AI hype with enterprise-grade precision. Kahol delivers a measured, rapid-fire breakdown of runtime architectures and real-world failure modes, balancing strategic vision with granular technical mechanics like saga orchestration and typed knowledge graphs. It feels like an executive masterclass on production-readiness, leaving the audience with a crisp, no-nonsense roadmap for deploying autonomous AI safely inside complex corporate environments.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) The future of managed services is shifting from billing for human labor supplemented by AI to operating and overseeing fleets of governed AI agents in true symbiosis. - [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) The Kagen architecture is built upon three foundational pillars: Kira (agent execution harness), Kagen (governance and orchestration platform), and AI in a Box (on-premises sovereign appliance). - [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) Governance, memory, orchestration, and cost controls cannot rely on project-level good intentions—they must be enforced as hard runtime properties. - [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) Early enterprise generative AI initiatives consistently failed due to five structural gaps: amnesia, lack of rollback mechanisms, hallucinations, opacity, and ungoverned execution. - [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) Kagen overcomes amnesia using a three-tier memory architecture (working memory, long-term promotion, and semantic recall) to retain state across interactions. - [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) By grounding LLM outputs in a governed knowledge graph ontology over enterprise data, hallucination rates drop to approximately 0.02%, compared to an industry average of roughly 30%. - [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) Saga orchestration provides checkpointed workflows with compensating rollback transactions and pause states at mandatory human approval gates. - [03:48](https://www.youtube.com/watch?v=SNrfmon1dnE&t=228s) The full-stack platform spans 20+ ingress channels (Teams, Slack, voice, IDEs), over 70 tools in Kira, 61 enterprise connectors, and integrations with 18+ LLM providers. - [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) Agents execute through a continuous five-stage PR runtime lifecycle: Perceive, Interpret, Infer, Execute, and Reflect. - [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) Dynamic model routing and governance reduce operational development task costs dramatically, benchmarked from $600 down to $10. - [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) Complex and high-risk scenarios leverage Swarm Debates—specialized persona committees that argue opposing perspectives until reaching verified convergence. - [07:49](https://www.youtube.com/watch?v=SNrfmon1dnE&t=469s) AI in a Box utilizes NVIDIA DGX hardware to run models up to 760 billion parameters on-premises for strict data sovereignty and local digital twin operations. - [07:49](https://www.youtube.com/watch?v=SNrfmon1dnE&t=469s) Enterprise adoption follows a structured three-phase roadmap: Land the stack (Months 0–2), Scale workflows (Months 2–6), and Run autonomous fleets (Month 6+).

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) The Shift to Symbiosis and Governed Agent Fleets

The conversation around artificial intelligence in enterprise environments is often dominated by two extremes: the fear that AI will outright replace human workers and business processes, or the dismissive view that it is merely "fancy autocomplete." In reality, the future of managed services lies in true symbiosis—deploying fleets of governed AI agents that carry out daily operational tasks while human experts supervise critical junctures. This approach centers on three core pillars: **Kira** (the agent harness where execution takes place), **Kagen** (the platform layer providing governance, orchestration, and guardrails), and **AI in a Box** (a self-contained, on-premises appliance for workloads that cannot leave an enterprise's physical or sovereign boundary).

When the initial wave of generative AI emerged, organizations relied on simple prompting. While proof-of-concept demos looked impressive, deployments consistently failed when exposed to production environments. An agentic enterprise cannot rely on good intentions from project teams or the assumption that every developer writes flawless code; governance must be a hard runtime property. The future winners in managed services will not sell billing hours assisted by AI—they will operate fleets of governed agents deployed across cloud and on-premises environments.

```mermaid flowchart TD accTitle: Three Pillars of the Kagen Symbiosis Architecture subgraph S1["1. Access & Execution"] KIRA["Kira Agent Harness<br/>(Every user touchpoint)"] end subgraph S2["2. Control & Governance"] KAGEN["Kagen Operating System<br/>(Runtime governance & memory)"] end subgraph S3["3. Deployment Options"] CLOUD["Managed Cloud"] BOX["AI in a Box<br/>(On-Premises / DGX Spark)"] end

KIRA --> KAGEN KAGEN --> CLOUD KAGEN --> BOX ```

- **Thesis for Managed Services:** Future leaders will operate and oversee fleets of governed agents rather than billing for human hours supplemented by AI. - **Runtime-Enforced Guardrails:** Memory, orchestration, human gates, and cost controls cannot be optional project scaffolding; they must be baked into the underlying operating system.

---

[02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) Five Structural Gaps of the First AI Wave

Early enterprise AI initiatives faltered due to five recurring structural gaps. These gaps serve as a practical evaluation checklist for any enterprise assessing AI platforms:

1. **Amnesia (It Forgets):** Standard LLM implementations lack state retention across interactions. Kagen resolves this through a three-tier memory architecture—working memory, promotion into long-term memory, and semantic recall—allowing agents to genuinely learn across user sessions. 2. **Lack of Rollback (It Can't Decide or Roll Back):** Agents frequently take uncoordinated actions without recovery mechanisms. Kagen uses saga orchestration with checkpointed workflows that execute compensating transactions on failure and enforce pause states at human approval gates. 3. **Hallucinations (It Hallucinates):** Unbounded generative models frequently produce invalid facts. By grounding all responses in a governed, typed knowledge graph ontology over enterprise data, hallucination rates drop to approximately **0.02%** (compared to an industry average hovering around 30%). Answers explicitly cite verified sources or explain why information is unavailable. 4. **Opacity (It's Invisible):** Traditional AI pipelines operate as black boxes. Kagen traces, indexes, and calculates token and computational costs for every single inference step and decision. 5. **Ungoverned Execution:** Ungoverned writes present substantial compliance risks. The platform enforces write-offs by default, requiring explicit, runtime-enforced sign-offs before external updates or mutations occur.

Enterprises that bypass this architectural scaffolding inevitably end up rebuilding isolated, custom implementations across different internal teams, destroying code reuse and auditability.

---

[03:48](https://www.youtube.com/watch?v=SNrfmon1dnE&t=228s) The Full-Stack Architecture: Kira and Kagen

The overall system spans from end-user input channels down to legacy backend enterprise infrastructure, structured in clean, domain-agnostic layers:

```mermaid flowchart TD accTitle: End-to-End Platform Stack Architecture subgraph IN["1. Ingress & Channels"] IO["20+ Messaging Platforms, Chat, Voice, IDEs, Terminals, APIs"] end subgraph HARNESS["2. Execution Harness: Kira"] LOOP["Single Agent Loop & Session Search"] TOOLS["70+ Tools, Skills & Delegation Profiles"] end subgraph OS["3. Platform & Governance: Kagen"] ENGINE["Agent Lifecycle (PR Cycle) & Router"] MODULES["3-Tier Memory | Saga Orchestration | Governed Ontology | Swarm Debates"] end subgraph BACKEND["4. Enterprise Connectors & Models"] CON["61+ Connectors (ERP, CRM, ITSM, Mainframes)"] LLMS["18+ Integrated LLM Providers & On-Prem Local Models"] end

IO --> HARNESS HARNESS --> OS OS --> BACKEND ```

- **Top Layer (I/O & Channels):** Work enters across 20+ channels including Microsoft Teams, Slack, voice telephony, developer IDEs, terminals, and raw APIs. - **Kira Harness Layer:** A unified agent execution loop equipped with over 70 tools, skill delegation routines, and searchable sessions. Kira allows users to interact with identical governed agents regardless of channel, eliminating per-channel custom rebuilds. - **Kagen Core Platform:** The engine managing the agent lifecycle, multi-tier memory, dynamic model routing, saga compensations, ontologies, persona swarms, approval mechanics, and observability. - **Enterprise Integrations & Model Layer:** 61 governed connectors linking into ERPs, CRMs, ITSMs, data platforms, and legacy banking mainframes, alongside integration with 18+ LLM providers with automatic fallback capabilities. - **Key System Properties:** Ubiquitous touchpoints (including inbound/outbound phone-based AIOps), complete multi-tenancy for managed service providers, full visibility with human cancellation overrides, and cross-agent task orchestration.

---

[06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) The Runtime Lifecycle, Swarm Debates, and Cost Optimization

At runtime, agents operate through a continuous five-stage loop known as the **PR lifecycle**:

$$\text{Perceive} \longrightarrow \text{Interpret} \longrightarrow \text{Infer} \longrightarrow \text{Execute} \longrightarrow \text{Reflect}$$

During this loop, the agent ingests data, interprets context, infers the next action, executes tool calls, and reflects on outcomes, iterating until convergence on an accurate solution. Dynamic routing ensures the lowest-cost capable model is selected for each sub-task. In real-world enterprise development benchmarks, this intelligent routing and governance reduced task programming costs from **$600 down to $10**.

For complex or high-risk challenges (such as security architecture reviews), Kagen activates **Swarm Debates**—a committee of specialized AI personas that argue opposing positions to reach verified convergence. Human approvals are managed asynchronously across web portals, email, and SMS. Non-negotiable runtime rules dictate that write-offs remain default, customer-facing agents cannot bypass human gates, and non-approved models are permanently blocked at the model registry level.

---

[07:49](https://www.youtube.com/watch?v=SNrfmon1dnE&t=469s) Real-World Deployments, AI in a Box, and Enterprise Rollout

The architecture is backed by 29 implemented demonstrations, proofs-of-concept, and production systems:

- **Healthcare Revenue Cycles:** An autopilot pipeline classifies claim denials and drafts sourced, fully cited appeals grounded directly in payer policies and contract ontologies. - **Supply Chain Management:** An ETA drift war room detects delivery schedule slippages, orchestrates multi-system remediation plans using saga rollbacks, and pauses at critical gates for logistician approvals. - **AI in a Box:** A hardware-integrated appliance powered by NVIDIA DGX Spark systems (or DGX Spark swarms) capable of hosting local models up to **760 billion parameters**. Designed for organizations demanding strict data sovereignty, this box enables enterprises to run full digital twins and autonomous operations entirely inside their own tenant walls without external data egress.

```mermaid flowchart LR accTitle: Enterprise AI Symbiosis Rollout Roadmap M1["Months 0 - 2<br/><b>Land the Stack</b><br/>Deploy Kira, Kagen & Connectors"] M2["Months 2 - 6<br/><b>Scale Workflows</b><br/>Expand Validated Domain Tasks"] M3["Months 6+<br/><b>Run the Fleet</b><br/>Full Autonomous Symbiosis"]

M1 --> M2 M2 --> M3 ```

The day-to-day symbiotic operating model follows a clean loop: agents process incoming task queues autonomously; high-stakes decision points automatically pause at the runtime level; human reviewers inspect and refine drafts; and execution resumes once authorized. Enterprise adoption follows a structured three-phase roadmap: **Months 0–2** to land the core stack, **Months 2–6** to scale operational workflows, and **Month 6 and beyond** to operate governed agent fleets at enterprise scale.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) The future winners in managed services will not sell billing hours assisted by AI—they will operate fleets of governed agents deployed across cloud and on-premises environments. — Kanav Kahol

> [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) An agentic enterprise cannot rely on good intentions from project teams or the assumption that every developer writes flawless code; governance must be a hard runtime property. — Kanav Kahol

> [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) Memory, orchestration, human gates, and cost controls cannot be optional project scaffolding; they must be baked into the underlying operating system. — Kanav Kahol

> [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) Enterprises that bypass this architectural scaffolding inevitably end up rebuilding isolated, custom implementations across different internal teams, destroying code reuse and auditability. — Kanav Kahol

> [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) By grounding all responses in a governed, typed knowledge graph ontology over enterprise data, hallucination rates drop to approximately 0.02% (compared to an industry average hovering around 30%). — Kanav Kahol

> [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) In real-world enterprise development benchmarks, this intelligent routing and governance reduced task programming costs from $600 down to $10. — Kanav Kahol

> [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) Non-negotiable runtime rules dictate that write-offs remain default, customer-facing agents cannot bypass human gates, and non-approved models are permanently blocked at the model registry level. — Kanav Kahol

> [07:49](https://www.youtube.com/watch?v=SNrfmon1dnE&t=469s) Designed for organizations demanding strict data sovereignty, this box enables enterprises to run full digital twins and autonomous operations entirely inside their own tenant walls without external data egress. — Kanav Kahol

## Mentioned Kanav Kahol, Kagen.ai, Kira, Kagen, AI in a Box, NVIDIA DGX Spark, Microsoft Teams, Slack

## Video description Kanav Kahol - Chief AI Officer, Kagen.ai

AI is transforming managed services from a labor-driven model to an intelligence-driven one. As organizations embed AI into service delivery, the focus is shifting toward defining clear roles for human expertise, autonomous systems, and business stakeholders while maintaining accountability and operational excellence. This session will explore emerging frameworks for AI governance, performance measurement, and outcome-based service delivery, with a particular focus on maximizing ROI from AI investments. Attendees will learn how leading enterprises are optimizing model selection, controlling token costs, and creating sustainable operating models that balance automation, trust, and business value. The session will provide a forward-looking view of how AI and managed services will evolve together to create more efficient, resilient, and scalable enterprises.

[Source](https://www.youtube.com/watch?v=SNrfmon1dnE)

中文简报

Kagen 部署受治理 AI 集群替代 IT 计费工时 托管服务提供商若不从出售人力转向运营具备运行时强制护栏、大幅削减执行成本的受治理智能体集群，就将走向失败。 Kanav Kahol 提出了企业人机共生的架构蓝图，认为托管服务的未来在于运营受治理的智能体集群，而非按人工工时计费。他分析了早期企业级 AI 的五大核心失效模式——健忘、缺乏回滚机制、幻觉、不透明性以及失控执行，并展示了如何通过操作系统级的运行时强制护栏来解决这些问题。该演讲详细介绍了 Kira 执行框架、Kagen 治理引擎以及本地部署“AI in a Box”一体机组成的全栈架构，重点阐述了动态模型路由与多智能体群体辩论如何在确保确定性“人在回路”监管的同时大幅削减运营成本。 托管服务正从以 AI 辅助的人工工时计费模式，转向在人机共生架构下运营和监督受治理的 AI Agent 集群。 (0:04) 早期企业级生成式 AI 常因健忘与缺乏回滚等缺陷受挫；通过引入三层记忆架构并锚定企业知识图谱本体，系统可将幻觉率从行业平均约 30% 压降至 0.02%。 (2:19) 全栈平台结合了 Kira 运行框架与 Kagen 治理引擎，打通 20 多个交互通道与 61 个企业连接器，并在关键决策点引入 Saga 事务回滚与人工强制审批门禁。 (3:48) 针对严格的数据主权合规要求，由 NVIDIA DGX Spark 驱动的“AI in a Box”软硬一体机支持运行高达 7600 亿参数的本地模型，使企业无需外泄数据即可完成全流程自主运维。 (7:49)

Reading summary

Ai4 argues that the future of enterprise managed services lies in operating governed fleets of AI agents in symbiosis with human supervisors rather than billing for human labor or relying on simple prompting. To transition AI from proof-of-concept to production, organizations must embed governance, memory, and saga orchestration directly into the platform runtime. The Shift Toward Symbiotic Agent Operations (0:04) Enterprise AI adoption is frequently framed either as wholesale human replacement or as basic autocomplete. In contrast, Ai4 positions the future of managed services as an operational symbiosis where fleets of autonomous agents handle routine execution while human specialists oversee critical decision gates. For this model to succeed, governance cannot depend on developer discipline or project-level scaffolding; it must operate as an immutable property enforced at runtime. The operational architecture rests on three pillars: Kira, a unified execution harness deployed across user touchpoints; Kagen, an underlying operating system providing runtime governance, memory, and orchestration; and AI in a Box, an on-premises appliance for sovereign environments. Managed service providers adopting this model are expected to shift from selling billable hours to managing and guaranteeing governed agent fleets deployed across hybrid cloud and local environments. Addressing Five Structural Enterprise AI Gaps (2:19) Early generative AI deployments routinely faltered in production due to five structural deficiencies: amnesia across interactions, an inability to roll back failed actions, factual hallucinations, operational opacity, and ungoverned execution. To resolve context loss, Kagen incorporates a three-tier memory model comprising working memory, promotion into long-term memory, and semantic recall, enabling agents to retain state across sessions. Workflow failures are mitigated through saga orchestration, which executes checkpointed compensating transactions on failure and enforces pause states at human verification gates. To counter hallucinations, the architecture grounds all model queries in a typed enterprise knowledge graph ontology. Ai4 claims this grounding reduces hallucination rates to roughly 0.02%—compared to an estimated 30% industry baseline—while requiring agents to cite verified sources or explicitly flag missing data. The platform also traces compute and token costs per inference step, while enforcing default write-offs that prevent external database updates without explicit, runtime-validated sign-offs. Full-Stack Integration and the Runtime PR Cycle (3:48) The system architecture spans four distinct layers: ingress across more than 20 channels (including messaging platforms, voice telephony, developer IDEs, and APIs); the Kira execution harness containing over 70 tools and skills; the Kagen governance engine; and an enterprise integration layer linking 61 connectors (such as ERPs, CRMs, and mainframes) with 18 model providers featuring automated fallback. Agents execute through a five-stage loop known as the PR lifecycle: Perceive, Interpret, Infer, Execute, and Reflect. For high-risk decisions, such as security evaluations, Kagen initiates Swarm Debates among opposing agent personas to reach consensus. Dynamic model routing assigns sub-tasks to the most cost-effective capable model, which Ai4 reports reduced task programming costs from $600 to $10 in development benchmarks. Customer-facing agents cannot bypass human approval checkpoints, and non-approved models are systematically blocked at the registry level. Real-World Use Cases, Sovereignty, and Rollout (7:49) The architecture is supported by 29 demonstrations, proofs-of-concept, and production implementations. Notable applications include healthcare revenue cycle pipelines that classify claim denials and draft cited appeals against payer policy ontologies, as well as supply chain logistics war rooms that identify delivery drift, plan multi-system remediations via saga rollbacks, and pause for human logistical approval. For enterprises constrained by strict compliance or physical data boundaries, the stack deploys through AI in a Box, an appliance powered by NVIDIA DGX Spark hardware that hosts local models up to 760 billion parameters without external data egress. Ai4 outlines a three-phase enterprise adoption roadmap: months 0 to 2 to deploy the Kira harness, Kagen platform, and connectors; months 2 to 6 to scale validated domain workflows; and month 6 onward to transition to autonomous agent fleet management supported by asynchronous human reviews over web, email, or SMS. Ai4 positions runtime governance and saga orchestration as the necessary foundation for viable enterprise agent fleets, yet long-term success will depend on whether asynchronous human approval gates can keep pace with autonomous execution without creating severe operational bottlenecks.

精炼阅读摘要

演讲者提出，企业级托管服务的未来并非出售由人工智能辅助的人工作业工时，而是转向人机共生模式，即在严格的运行时治理下集中运营和监督自主AI智能体集群。 托管服务向共生架构与受管智能体集群演进 (0:04) 当前企业对人工智能的讨论常陷于两个极端：要么担忧AI将彻底取代人类员工与业务流程，要么将其贬低为高阶的自动补全工具。演讲指出，托管服务的真正未来在于人机共生模式，即部署受控的AI智能体集群处理日常运营任务，由人类专家把控关键决策节点。该架构由三大核心支柱支撑：负责在所有用户触点执行操作的智能体容器Kira、提供运行时治理与编排底座的操作系统Kagen，以及专为高数据主权与物理隔离需求设计的本地一体机AI in a Box。 在生成式AI初期，企业普遍依赖简单的提示词工程构建概念验证，但这类方案在实际生产环境中频繁失效。演讲者强调，代理型企业无法寄希望于项目团队的良好意愿或完美的编码能力，系统治理、记忆留存、人工审批关卡和成本控制必须直接内嵌在底层操作系统中，作为硬性的运行时属性强制执行。 第一代企业AI的五大结构性缺陷及治理方案 (2:19) 早期的企业AI项目受挫主要源于五大结构性缺陷。首先是健忘问题，传统大语言模型缺乏跨交互的状态保持能力，该架构通过工作记忆、长期记忆沉淀和语义检索组成的三层记忆系统解决，使智能体能够跨会话持续学习。其次是缺乏回滚机制，针对智能体盲目行动的风险，系统采用基于Saga编排的分步工作流，在故障时执行补偿事务，并在人工审批点强制暂停。第三是幻觉率高，通过将模型输出强行锚定在企业数据的受管类型化知识图谱本体上，据称幻觉率可从行业平均约30%骤降至0.02%，且要求智能体明确引用已验证来源或说明数据缺失。 此外，针对传统AI流程黑盒化导致的不可见性，平台对每次推理和决策步骤进行全量链路追踪、索引并核算Token与计算成本；针对未受管执行带来的合规风险，系统默认执行写入拦截机制，任何涉及外部系统变更的写操作都必须在运行时通过明确的授权审批。演讲者指出，绕过这些基础架构的企业往往会在各业务线重复造轮子，导致代码复用和合规审计彻底失效。 Kira与Kagen全栈架构与企业级集成能力 (3:48) 整个技术栈从前端触点延伸至传统后端基础设施，分为清晰的分层解耦设计。最上层为接入信道层，支持包括Microsoft Teams、Slack、语音电话、集成开发环境（IDE）、终端以及API等20余种交互途径。紧接着是Kira执行容器层，它具备统一的智能体运行闭环、会话检索能力以及包含70余种工具与技能委派配置的工具箱，使用户在不同渠道面对的智能体保持一致行为，消除了按渠道重复开发的问题。 下层则由Kagen核心平台支撑，负责管理智能体生命周期、多层记忆调度、动态模型路由、Saga补偿事务、知识本体以及可观测性。底层集成层则提供了61个预置的企业连接器，可直连ERP、CRM、ITSM及传统银行大型主机，并接入了18家以上的大语言模型提供商与本地模型，支持自动故障转移。该系统具备支持电话出入呼叫的AIOps能力、面向托管服务商的多租户架构，以及允许人工随时介入取消任务的全程可视化监控。 PR运行时生命周期、群体辩论与成本优化 (6:15) 智能体在运行阶段遵循被称为PR生命周期的五步闭环，即感知（Perceive）、解释（Interpret）、推断（Infer）、执行（Execute）和反思（Reflect）。智能体持续摄入数据、理解上下文、推断下一步行动并调用工具，通过反复反思迭代直至收敛于精确方案。系统还引入了动态模型路由机制，为各个细分任务匹配成本最低且能力达标的模型。演讲者引用实际基准数据称，该机制将特定开发任务的单次执行成本从600美元大幅削减至10美元。 对于安全架构审查等复杂或高风险任务，Kagen会触发群体辩论机制，组织多个具有特定角色的AI智能体委员会进行正反方辩论，以达成经严密论证的共识。高风险环节的人工审批则可通过网页、电子邮件或短信异步完成。系统设立了不可逾越的硬性规则：写操作默认禁用、面向客户的交互严禁绕过人工关卡、未经平台注册中心批准的模型在底层被彻底拦截阻断。 落地案例、本地化一体机与企业推进路线图 (7:49) 该架构目前已落地29项概念验证与生产系统。在医疗营收周期管理中，智能体自动分类拒赔单据，并依据付款方政策和合同本体生成带有来源引用的申诉材料；在供应链管理中，系统建立了预计到达时间（ETA）漂移作战室，负责检测延误并生成跨系统修复计划，利用Saga机制在关键物流审批节点自动暂停等待确认。针对数据主权和合规极高要求的场景，方案推出了AI in a Box一体机，基于NVIDIA DGX Spark系统或集群构建，可本地运行高达7600亿参数的大模型，在无外部数据外流的前提下支撑企业数字孪生与自主运营。 企业采纳这一共生模式通常遵循明确的三阶段推进路线图：前两个月为底座部署阶段，完成Kira、Kagen及连接器的落地；第二至第六个月为工作流扩展阶段，逐步拓展经业务验证的专业领域任务；第六个月之后则进入集群运营阶段，全面迈入规模化的人机协同共生运营。 演讲者所描绘的智能体运营模式将治理与安全约束视为不可妥协的前提条件，强调彻底摆脱单纯的人工工时计费逻辑。然而，该方案的高度自主化运转在现实中仍严格依赖高质量的企业知识图谱、完善的底层连接器以及关键节点上人类专家的持续监督与签字授权。

Detailed notes — zh

## Background 该内容由 Ai4 频道于 2026 年 8 月 17 日发布在 YouTube 平台上，演讲嘉宾为 Kagen.ai 的首席 AI 官 Kanav Kahol。视频简介指出，本次演讲主要探讨 AI 如何推动托管服务从劳动密集型向智能驱动型模式转型，并深入解析在确保可信度与业务价值的前提下，企业如何通过 AI 治理框架、模型路由优化和成本控制实现人机共生。 ## TL;DR Kanav Kahol 提出未来托管服务的核心在于运营受治理的自主 AI Agent 机队，而非简单的人工替代。他剖析了第一代大模型应用在记忆、回滚、幻觉和治理等方面的五大缺陷，并介绍了由 Kira 接入框架、Kagen 操作系统及 AI in a Box 硬件构成的完整解决方案。通过 PIIER 循环、Saga 编排、蜂群辩论机制以及受治理的企业本体知识图谱，该方案将幻觉率压低至 0.02% 并大幅削减 Token 成本，为企业提供了清晰的三阶段落地演进路线。 ## Atmosphere & style 演讲节奏紧凑干练且极具工程实战感，Kanav 的语气沉稳自信、逻辑严密，摒弃了空洞的行业概念炒作，直击大模型落地生产环境时的痛点与结构性缺陷。整场分享技术架构清晰明了，通过展示极具视觉冲击力的数据对比与落地方案，营造出一种技术已完全就绪、变革蓄势待发的务实与紧迫氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) 托管服务的未来范式是人机共生（Symbiosis），核心不再是售卖 AI 辅助下的人工工时，而是运营和管理受治理的自主 Agent 机队。 - [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) Kagen 平台通过 Kira 运行框架、Kagen 操作系统及 AI in a Box 硬件方案三大支柱构建完整的人机共生技术架构。 - [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) 第一代 AI 的结构性缺陷在于容易遗忘、无法自主决断回滚、幻觉严重、执行过程黑盒以及缺乏治理。 - [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) 通过引入工作记忆、长期记忆和语义召回的三层记忆系统，Agent 得以实现跨会话持续学习。 - [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) 将答案严格锚定在受治理的企业本体知识图谱（Governed Ontology）中，成功将模型幻觉率从约 30% 降至 0.02%。 - [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) 采用带人类审批门禁（Human Gates）的 Saga 编排工作流，支持在失败时自动执行补偿回滚并强制人工确认关键步骤。 - [03:48](https://www.youtube.com/watch?v=SNrfmon1dnE&t=228s) Kira 交互前端支持 Teams、Slack、语音通话等多渠道接入，并具备可观测、随时人工中断及 AI 编排 AI 等能力。 - [03:48](https://www.youtube.com/watch?v=SNrfmon1dnE&t=228s) 系统集成了 61 个连接器以适配 ERP/CRM/主机等系统，并支持跨 18 种以上大模型智能调用与故障回退。 - [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) Agent 运行时遵循感知（Perceive）、解析（Interpret）、推理（Infer）、执行（Execute）与反思（Reflect）的 PIIER 闭环生命周期。 - [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) 通过智能模型路由器动态派发低成本模型，可将编程等复杂任务的单次执行成本从 600 美元降至 10 美元。 - [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) 对于高风险及复杂决策，引入 Agent 蜂群辩论机制（Swarm Debate）由多个 AI 角色相互辩论直至达成共识。 - [07:49](https://www.youtube.com/watch?v=SNrfmon1dnE&t=469s) 针对严苛的数据主权与本地化部署需求，AI in a Box 硬件一体机可在本地运行高达 7600 亿参数的大规模模型。 - [07:49](https://www.youtube.com/watch?v=SNrfmon1dnE&t=469s) 企业实施人机共生应分阶段推进：0-6 周落地基础架构，2-6 个月扩展验证工作流，6 个月后全面运营 Agent 机队。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) 共生而非替代：受治理的 Agent 机队与托管服务的未来 来自 Kagen.ai 的首席 AI 官 Kanav Kahol 开门见山地指出，AI 的出现并不是为了取代任何员工或流程，也不是单纯让人们把 AI 当作一种“花哨的自动补全工具（fancy autocomplete）”。他坚信未来的核心范式是**人机共生（Symbiosis）**——由一套受治理的 AI Agent 机队去执行具体工作，而在关键节点由人类进行监督与把关。托管服务（Managed Services）未来的赢家，绝不再是靠售卖“AI 辅助下的人工工时”，而是去运营和管理一整支受治理的自主 Agent 机队。 为了将这一愿景落地，Kanav 概述了核心技术支撑的三大支柱： 1. **Kira（Agent 运行框架/Harness）**：Agent 实际执行工作的交互前端与运行环境。 2. **Kagen（Agent 平台/操作系统）**：实现任务编排、模型路由及运行治理的平台中枢。 3. **AI in a Box（一体机硬件方案）**：专为数据主权和严苛本地化部署需求打造的私有化运行载体。 ```mermaid flowchart TD accTitle: Kagen AI Platform Architecture subgraph IO["接入与交互层 (Touchpoints & Kira Harness)"] A["多渠道接入: Teams / Slack / 语音电话 / IDE / API"] B["Kira Harness: 单一 Agent 循环, 70+ 工具库, 会话检索"] A --> B end subgraph OS["平台治理与编排核心 (Kagen OS Runtime)"] C["生命周期管理 (PIIER 循环)"] D["三层记忆系统 (工作记忆 / 长期记忆 / 语义召回)"] E["Saga 编排与人类审批门禁 (Human Gates)"] F["受治理知识图谱本体 (Governed Ontology)"] G["模型智能路由 (Model Router) 与 FinOps 成本控制"] B --> C C --> D C --> E C --> F C --> G end subgraph Backend["底层连接与基础设施 (Integrations & Hardware)"] H["61+ 治理连接器: ERP / CRM / ITSM / 主机 Mainframe"] I["18+ LLM 供应商模型适配与回退"] J["部署形态: 云端托管 / AI in a Box (NVIDIA DGX 本地私有化)"] E --> H G --> I C --> J end ``` --- ### [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) 第一代 AI 失败的五大结构性缺陷与应对方案 回顾大模型落地的第一波浪潮，Kanav 提到当时大家都在写 Prompt，Demo 阶段看似惊艳，但一旦进入生产环境就会全面崩溃。很多企业在没有系统脚手架的情况下盲目开发，导致每个团队都在重复造轮子，既无法复用也无法审计。针对第一代 AI 的五大结构性缺陷，Kagen 提出了对应的全生命周期运行时解决方案，可作为企业评估 AI 供应商的标准清单： - **容易遗忘（It forgets）**：构建了**三层记忆系统**（工作记忆、晋升至长期记忆、语义召回），使 Agent 能够真正跨会话持续学习。 - **无法自主决断或回滚（It can't decide or roll back）**：引入带有人类把关门禁（Human Gates）的 **Saga 编排机制**，设置检查点工作流，失败时自动执行补偿回滚，并在关键节点强制暂停等待人工确认。 - **幻觉问题（It hallucinates）**：将答案严格锚定（Grounding）在**受治理的企业本体知识图谱（Governed Ontology）**中。该知识图谱基于企业实际数据构建且强类型化，回答必须引用来源；若无法回答则明确解释原因。Kagen 将业界的幻觉率从通常的约 30% 骤降至 **0.02%**。 - **过程不可见（It's invisible）**：每一个推理步骤与工具调用均实现可追踪、可搜索，并实时核算 Token 成本。 - **缺乏治理（Ungoverned）**：默认关闭高危写操作（Write-offs are by default），所有审批均在底层运行时严格执行，从根源上杜绝配置绕过。 --- ### [03:48](https://www.youtube.com/watch?v=SNrfmon1dnE&t=228s) Kagen 系统全景与 Kira 无缝接入体验 从整体系统架构来看，自上而下涵盖了接入层、运行层与底层集成： 1. **I/O 与全渠道接入**：覆盖 Teams、Slack、语音通话、IDE、API 等 20 多种主流平台。Kira 强调“在员工已有的工作阵地与其相遇”，无需为每个渠道单独重建逻辑。 2. **Kira 运行框架**：内置统一的 Agent 核心循环与 70 多种现成工具技能，支持任务委派与会话追踪。 3. **Kagen 平台治理层**：统管 Agent 生命周期、记忆、模型路由、Saga 编排、本体图谱、Agent 蜂群辩论（Swarm Debate）、人工审批及可观测性。 4. **底层集成与模型层**：提供针对 ERP、CRM、ITSM、数据平台乃至老旧大型机（Mainframe）的 **61 个受治理连接器**，并支持跨 **18 种以上主流大模型**的智能调用与故障回退。整套体系完全领域无关（Domain Agnostic），适用于各类行业垂直场景。 Kira 具备四项核心特征： - **无处不在（Ubiquitous）**：同一个受治理 Agent 可以在聊天软件、IDE 中运行，甚至能直接拨打语音电话进行 AIOps 运维。 - **可观测与可中断（Observable & Interruptible）**：所有工具调用对人类完全透明，且支持随时人工中断取消。 - **多租户隔离（Multi-tenant）**：支持隔离配置并发运行，托管服务商可用单套部署服务多个客户。 - **AI 编排 AI（Orchestrates other AI）**：可将复杂阶段委派给专门的子 Agent 并在完成后审查其产出。 --- ### [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) PIIER 生命周期、蜂群辩论与硬性治理防线 在运行时治理层面，Agent 遵循严格的 **PIIER 五阶段生命周期**： - **Perceive（感知）**：接收并获取外部输入与上下文数据。 - **Interpret（解析）**：解析数据意图与结构。 - **Infer（推理）**：推断下一步所需执行的操作。 - **Execute（执行）**：调用工具并执行对应任务。 - **Reflect（反思）**：对执行结果进行自省评估，持续循环迭代直至收敛出正确结果。 在成本与安全控制上，系统采用**智能模型路由器**动态分派能够胜任任务的最廉价模型。在现场展示的 Kagen AD 平台中，某个编程任务的单次执行成本成功从 **$600 降低至 $10**。针对复杂安全或高风险决策，系统引入 **Agent 蜂群辩论机制（Swarm Debate）**——由多个 AI 角色委员会针对同一问题展开激烈辩论，直至达成共识。 Kanav 特别强调了平台的**不可妥协底线（Non-negotiables）**：面向客户的 Agent 严禁绕过人工审批门禁，高危写入操作默认阻断，未经审批的模型在注册表中直接被拦截。 --- ### [07:49](https://www.youtube.com/watch?v=SNrfmon1dnE&t=469s) 落地实践、AI in a Box 硬件方案与演进路线 Kagen 目前已落地 29 个覆盖 PoC、演示及真实生产环境的项目： - **医疗收入周期管理**：打造拒赔与申诉自动驾驶仪（Denials and Appeals Autopilot），自动分类拒赔原因，基于支付方知识本体撰写带有精准引用的申诉材料。 - **供应链管理**：构建“ETA 偏差作战室（ETA Drift War Room）”，实时检测物流预计到达时间偏差，自动编排带回滚机制的补救措施，并在关键节点等待人工审批。 针对数据主权要求极高、数据严禁出内网的政企客户，Kagen 推出了实体硬件产品 **AI in a Box**。该设备基于 NVIDIA DGX Spark 或其集群构建，能在本地一体机中运行高达 **7600 亿（760B）参数**的大规模模型，为企业搭建全套业务运行的数字孪生环境，也适用于中小企业本地一键运行整套 Agent 平台。 企业迈向人机共生的实施路线清晰明了： - **第 0 至 6 周**：落地部署基础技术栈（Land the stack）。 - **第 2 至 6 个月**：验证并扩展有效的工作流规模（Scale that works）。 - **第 6 个月及以后**：全面常态化运营受治理的 Agent 机队（Run the fleet）。 Kanav 总结道，人机共生的模式与平台已经完全就绪，企业唯一需要做的决定，就是挑选出第一个准备改造的工作流。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) AI 的出现并不是为了取代任何员工或流程，也不是单纯让人们把 AI 当作一种“花哨的自动补全工具（fancy autocomplete）”。 — Kanav Kahol > [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) 未来的核心范式是人机共生（Symbiosis）——由一套受治理的 AI Agent 机队去执行具体工作，而在关键节点由人类进行监督与把关。 — Kanav Kahol > [00:04](https://www.youtube.com/watch?v=SNrfmon1dnE&t=4s) 托管服务（Managed Services）未来的赢家，绝不再是靠售卖“AI 辅助下的人工工时”，而是去运营和管理一整支受治理的自主 Agent 机队。 — Kanav Kahol > [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) 当时大家都在写 Prompt，Demo 阶段看似惊艳，但一旦进入生产环境就会全面崩溃。 — Kanav Kahol > [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) Kagen 将业界的幻觉率从通常的约 30% 骤降至 0.02%。 — Kanav Kahol > [02:19](https://www.youtube.com/watch?v=SNrfmon1dnE&t=139s) 默认关闭高危写操作，所有审批均在底层运行时严格执行，从根源上杜绝配置绕过。 — Kanav Kahol > [03:48](https://www.youtube.com/watch?v=SNrfmon1dnE&t=228s) 在员工已有的工作阵地与其相遇，无需为每个渠道单独重建逻辑。 — Kanav Kahol > [06:15](https://www.youtube.com/watch?v=SNrfmon1dnE&t=375s) 面向客户的 Agent 严禁绕过人工审批门禁，高危写入操作默认阻断，未经审批的模型在注册表中直接被拦截。 — Kanav Kahol > [07:49](https://www.youtube.com/watch?v=SNrfmon1dnE&t=469s) 人机共生的模式与平台已经完全就绪，企业唯一需要做的决定，就是挑选出第一个准备改造的工作流。 — Kanav Kahol ## Mentioned Kanav Kahol, Kagen.ai, Kira, Kagen, AI in a Box, NVIDIA DGX, Microsoft Teams, Slack, PIIER, Saga, ITSM, ERP, CRM, ETA Drift War Room, Denials and Appeals Autopilot ## Video description Kanav Kahol - Chief AI Officer, Kagen.ai AI is transforming managed services from a labor-driven model to an intelligence-driven one. As organizations embed AI into service delivery, the focus is shifting toward defining clear roles for human expertise, autonomous systems, and business stakeholders while maintaining accountability and operational excellence. This session will explore emerging frameworks for AI governance, performance measurement, and outcome-based service delivery, with a particular focus on maximizing ROI from AI investments. Attendees will learn how leading enterprises are optimizing model selection, controlling token costs, and creating sustainable operating models that balance automation, trust, and business value. The session will provide a forward-looking view of how AI and managed services will evolve together to create more efficient, resilient, and scalable enterprises. [Source](https://www.youtube.com/watch?v=SNrfmon1dnE)

Source quotes

[0:04] Kanav Kahol: AI 的出现并不是为了取代任何员工或流程，也不是单纯让人们把 AI 当作一种“花哨的自动补全工具（fancy autocomplete）”。 [0:04] Kanav Kahol: 未来的核心范式是人机共生（Symbiosis）——由一套受治理的 AI Agent 机队去执行具体工作，而在关键节点由人类进行监督与把关。 [0:04] Kanav Kahol: 托管服务（Managed Services）未来的赢家，绝不再是靠售卖“AI 辅助下的人工工时”，而是去运营和管理一整支受治理的自主 Agent 机队。 [2:19] Kanav Kahol: 当时大家都在写 Prompt，Demo 阶段看似惊艳，但一旦进入生产环境就会全面崩溃。 [2:19] Kanav Kahol: Kagen 将业界的幻觉率从通常的约 30% 骤降至 0.02%。 [2:19] Kanav Kahol: 默认关闭高危写操作，所有审批均在底层运行时严格执行，从根源上杜绝配置绕过。 [3:48] Kanav Kahol: 在员工已有的工作阵地与其相遇，无需为每个渠道单独重建逻辑。 [6:15] Kanav Kahol: 面向客户的 Agent 严禁绕过人工审批门禁，高危写入操作默认阻断，未经审批的模型在注册表中直接被拦截。 [7:49] Kanav Kahol: 人机共生的模式与平台已经完全就绪，企业唯一需要做的决定，就是挑选出第一个准备改造的工作流。

Source quotes

[0:04] Kanav Kahol: The future winners in managed services will not sell billing hours assisted by AI—they will operate fleets of governed agents deployed across cloud and on-premises environments. [0:04] Kanav Kahol: An agentic enterprise cannot rely on good intentions from project teams or the assumption that every developer writes flawless code; governance must be a hard runtime property. [0:04] Kanav Kahol: Memory, orchestration, human gates, and cost controls cannot be optional project scaffolding; they must be baked into the underlying operating system. [2:19] Kanav Kahol: Enterprises that bypass this architectural scaffolding inevitably end up rebuilding isolated, custom implementations across different internal teams, destroying code reuse and auditability. [2:19] Kanav Kahol: By grounding all responses in a governed, typed knowledge graph ontology over enterprise data, hallucination rates drop to approximately 0.02% (compared to an industry average hovering around 30%). [6:15] Kanav Kahol: In real-world enterprise development benchmarks, this intelligent routing and governance reduced task programming costs from $600 down to $10. [6:15] Kanav Kahol: Non-negotiable runtime rules dictate that write-offs remain default, customer-facing agents cannot bypass human gates, and non-approved models are permanently blocked at the model registry level. [7:49] Kanav Kahol: Designed for organizations demanding strict data sovereignty, this box enables enterprises to run full digital twins and autonomous operations entirely inside their own tenant walls without external data egress.

Original transcript

[0:04] Good afternoon, everyone. I'm here on behalf  of Kagan AI and I want to spend the next 10 [0:09] minutes or maybe nine on the future of managed  services and the fact that I don't think AI is [0:15] replacing any people or any processes and it's  not just people using AI as a fancy autocomplete. [0:23] I do believe it's symbiosis. Fleet of governed AI  agents doing the work with people supervising the [0:29] work when it actually matters. To make that real,  I'm going to talk about three different things. [0:35] Kira, our agent harness, that's where the work  happens. Kagan, which is actually our agentic [0:41] platform and how the work is governed and then  AI in a box when the work that you are doing has [0:48] to stay within inside your walls. So, let me walk  you through this and how it matters. Essentially, [0:53] we know that when the first wave of AI actually  broke that came in, we were all essentially just [0:59] prompting stuff. Demos really worked beautifully.  But then we actually had the situation where [1:05] everything was breaking. We know that for  example, when this was actually happening, [1:10] things were breaking in production. Second,  what the agentic enterprise actually needs is [1:14] an operating system. You need durable memory,  you need orchestration, human approval gates, [1:21] grounded answers, observability and cost control.  And essentially, you will see a bunch of solutions [1:26] across each one of them. Kagan actually  brings all of this together and critically, [1:31] those have to be properties of runtime. Right?  Not just the good intentions of a project team. [1:36] I often joke if all of us had great project  teams and everybody was a great developer, [1:41] we won't need any of this. But the problem is not  all of us are great all the time. And that's where [1:46] actually you need Kagan to come in and really  work on this. Third, that operating system [1:51] that exists today, Kira actually puts agents on  every touchpoint your people already use. So, [1:57] Kagan governs every action underneath, and you  can run it as a managed services in the cloud or [2:03] entirely on premise. This is very very important  that you can actually run all of this on premise, [2:08] and the takeaway line at the bottom is the whole  thesis. The winners in the managed services [2:12] won't sell hours assisted by AI, they'll  actually operate fleets of governed agents. [2:19] Why did the first wave fail? Well, there were  five structural gaps, and I'd encourage you [2:24] to read this list as a checklist for any AI vendor  you actually evaluate, right? It forgets. We know [2:30] that we need to actually go ahead and look at  that. So, we built a three-tier memory system, [2:36] working memory that gets promoted into long-term  memory and semantic recall. Agents that genuinely [2:42] learn across sessions. It can't decide or roll  back. So, we built saga orchestration with human [2:49] gates, checkpointed workflows that compensate  on failure, and pause for every real sign-off. [2:55] It hallucinates. So, we ground our answers into  a governed ontology. Please come and visit us [3:00] on our booth number 350. You'll actually see  these ontologies. We've been told it's the [3:04] best-looking ontology. Um we're quite proud  of it. A typed knowledge graph that works [3:09] over your data where answers cite sources and  explains why they can't. Our hallucination rates [3:14] are around 0.02% compared to industry which is  generally seeing around 30%. It's invisible. [3:21] So, every inference step and decision is  traced, searchable, and costed in our system. [3:26] And we also know that lot of AI up till now  has been ungoverned. So, with us, write-offs [3:32] are by default, and approvals are enforced in  the runtime. One sentence to remember here, [3:37] every enterprise that skipped this  scaffolding ended up rebuilding it [3:41] slightly differently per team with no reuse  and no re-audit. That's what Kagan avoids. [3:48] So, we've got essentially the whole system  on one page, read it from top to bottom. [3:54] At the top, I/O people and channel, your  chat, your 20-plus messaging platforms like [3:59] Teams, Slacks, voice, IDE, APIs, that's where the  work arrives. Here are the harness underneath, [4:06] one agent loop, 70-plus tools, skills,  delegation, sessions you can actually search. [4:14] It meets people where they already work. No per  channel rebuild. This is something that we're [4:19] quite passionate about. Below that, the Kegan,  the platform library, that's where the agent life [4:24] cycle, memory, model routing, sagas, ontology,  swarms, approval, and observability states. [4:31] It makes every action governable and  configuration cannot bypass it. This is again [4:37] very, very important. And at the bottom,  the two things every enterprise cares about, [4:41] your system, 61 govern connectors into ERP,  CRM, ITSMs, data platforms, even mainframe, [4:49] and your models. We We work with a lot of banks  that are actually still on mainframes. So, we're [4:53] actually using that as a way of going ahead and  providing connectors for everything. Both layers [4:58] are deliberately domain agnostic. Nothing here  is fed to the vertical. So, everything is what [5:04] actually matters is is covered here. The Kegan  architecture. So, we are actually going to be [5:11] Let's zoom into Kegan essentially for a moment.  We're actually looking at the fact that the edges [5:16] you see are the entry points, the terminal, the  messaging gateway, the 20-plus platform adapters, [5:22] IT integration and schedule jobs. In the middle,  one agent core essentially serves those surfaces, [5:28] prompts, context engineering, provider resolution  across 18-plus LLM tools with fallbacks, [5:33] and tool dispatches with 70-plus tools. The  four properties I want you to take away it's [5:38] ubiquitous, the same governed agent answers in  Teams, on the phone, in the IDE. We can actually [5:44] make phone calls with our system, and your AI ops,  for example can happen exactly on a phone call. [5:49] It's observable and interruptible. Every tool  call is visible and cancelable by human agents. [5:56] It's multi-tenant isolated profiles that run  concurrently, which for a managed service [6:01] business means one deployment many clients and  it orchestrates other AI. It delegates phases [6:07] to specialized workers and reviews their  outputs. Symbiosis is inside the machine too. [6:15] The Kegan architecture, that's where we are  actually looking at governance as being a property [6:20] of the real-time. What we are at of the run-time  rather. Follow the four top connectors and just [6:26] real data. Agents run five stage life cycle. This  is called the PR life cycle, which is perceive, [6:33] interpret, inferred, execute, reflect. Right? So  it actually gets the data, that's the perceiving [6:39] part. It interprets the data. It infers what it  needs to do with it. It then will execute and then [6:45] it will finally reflect on its decision and redo  this cycle till it actually gets and converges [6:51] to the right answer. A router is used to make  sure that we're picking up the cheapest capable [6:56] model for each task. Sagas orchestrate, humans  approve. Every step is traced and costed. Come [7:02] and see our platform called Kegan ad where I can  actually tell you what was the programming cost [7:08] for each one of the aspects and how we actually  reduced it from $600 to $10. The eight modules [7:14] below are the toolbox, memory, routing, ontology,  swarm debate. So if you're dealing with security, [7:20] you'll have swarms that likely debate over it and  give you the answers on that particular side. It's [7:25] a committee of AI personas that argue hard over  a problem to convergence. Human approvals to [7:30] manage portals over emails, SMSs and of course  full finops directly work. The banner at the [7:36] bottom is what I'd call as our non-negotiables.  Human approval, customer facing agents cannot be [7:41] configured around it. Write-offs are by default  non-approved models are blocked at the registry. [7:49] The proof actually lies in the fact that  we have been building a variety of these [7:53] solutions. So, right there, we have around  29 implemented demonstrations as well as POCs [7:59] and real production projects. Um let's talk about,  for example, the healthcare revenue cycles, the [8:04] denials and appeals autopilot directly classifies  denials, draft cited appeals grounded in payers, [8:10] and allows the ontology to do that. In supply  chain, the ETA drift war room can actually [8:15] detect the drift, orchestrate remediations with  rollback and pauses for approvals, and so on. [8:22] We also have a solution which is called AI in a  box. This is an actual product, not an AI image, [8:27] um where we actually have the ability to guide an  NVIDIA DGX Spark or a swarm of NVIDIA DGX Sparks. [8:35] We can actually host models as high as 760 billion  parameters in this little box. And then there are [8:41] specialized skills for a variety of operations  right from your operations, IT, sales, marketing, [8:49] allowing companies and enterprises to create a  full digital twin, and actually go ahead and play [8:54] around with their entire environment. But also for  small and medium businesses to actually let this [9:01] run your entire agentic platform. This is  something that we have. We are working with [9:05] companies that do not that really believe in  data sovereignty and don't want their data [9:10] to leave their tenant. The operating model, what  does the symbiosis actually look like day-to-day? [9:18] Five steps on a loop, agents execute the queue,  high-stakes steps high-stakes steps pause, [9:24] the runtime holds them, person reviews,  edits the draft, and so on. And finally, [9:29] let's talk a little bit about what what is the  path forward. What we actually do is we zero to [9:35] six, we land the stack, month 2 to 6, scale  that actually works. And month 6 and beyond, [9:41] we run the fleet. So, symbiosis is the model.  The platform exists. The only decision left [9:47] for you is the first workflow. And I'd love to  work with you. Please visit us on booth 350. [9:52] I'm Kanav, and I'm happy to take any more  questions that you have at the booth. Thank you.