deanlu.ai

Core AI Platforms & Agents

8 min read

Source Snapshot

  • Origin: NVIDIA product pages, docs, Newsroom, and technical blog materials
  • Type: Research synthesis
  • Author / org: NVIDIA
  • One-line takeaway: NVIDIA’s agent strategy is about making autonomous agents deployable, observable, secure, and self-hostable.

Garden Card

This note maps NVIDIA’s core enterprise agent stack: NeMo Agent Toolkit, NIM, AI-Q, NemoClaw, and OpenShell.

这篇笔记梳理 NVIDIA 的企业智能体核心栈：NeMo Agent Toolkit、NIM、AI-Q、NemoClaw 和 OpenShell。

  • Core question: What stack is needed to run agents safely inside enterprise infrastructure? 核心问题：在企业基础设施中安全运行智能体需要什么技术栈？

  • Operational value: It separates model serving, workflow design, private data retrieval, runtime policy, and observability. 运营价值：它区分模型服务、工作流设计、私有数据检索、运行时策略和可观测性。

  • Best connection: Hardware Architecture & Computing Infrastructure, Open Models & Industry Verticals, NVIDIA Factory Operations Blueprint FOX 最适合连接的内容：硬件基础设施、开放模型/行业垂直和 FOX。

1. Executive Summary

NVIDIA is packaging agentic AI as a layered enterprise stack. NIM serves models, NeMo Agent Toolkit builds and optimizes workflows, AI-Q demonstrates enterprise research agents, OpenShell enforces runtime policy, and NemoClaw packages always-on agent patterns.

NVIDIA 正把智能体人工智能封装成分层企业栈。NIM 提供模型服务，NeMo Agent Toolkit 构建和优化工作流，AI-Q 展示企业研究智能体，OpenShell 执行运行时策略，NemoClaw 封装常驻智能体模式。

The strategic shift is that agent runtime security and model serving become infrastructure concerns rather than application prompt concerns.

战略变化在于，智能体运行时安全和模型服务变成基础设施问题，而不只是应用提示词问题。

  • Main idea: Agent platforms need separate control planes. 主要观点：智能体平台需要独立控制平面。

  • Why now: Agents increasingly read files, call APIs, use credentials, and run continuously. 为什么现在重要：智能体越来越多地读取文件、调用 API、使用凭证并持续运行。

  • Where it applies: Private research agents, manufacturing assistants, engineering copilots, workflow agents, and local AI services. 可以应用的场景：私有研究智能体、制造助手、工程副驾、工作流智能体和本地 AI 服务。

Decision Signal

Treat agent runtime security and model serving as first-class infrastructure, not application-level prompts.

2. Key Technical Terms

Use these terms to describe NVIDIA’s agent platform stack.

这些术语可以描述 NVIDIA 的智能体平台栈。

  • NeMo Agent Toolkit / NeMo 智能体工具包: Framework-agnostic layer for connecting, profiling, evaluating, and optimizing agents.

    用于连接、画像、评估和优化智能体的框架无关层。

  • NIM / 推理微服务: Optimized production inference microservices for foundation models.

    面向基础模型的优化生产推理微服务。

  • AI-Q Blueprint / AI-Q 蓝图: Reference architecture for enterprise research agents over private and external data.

    面向企业研究智能体的参考架构。

  • NemoClaw / NemoClaw 参考栈: Reference stack for OpenClaw-style always-on agents with privacy controls.

    面向 OpenClaw 风格常驻智能体并带隐私控制的参考栈。

  • OpenShell / 安全运行时: Sandboxed runtime that enforces filesystem, network, credential, and inference policy.

    执行文件系统、网络、凭证和推理策略的沙箱运行时。

3. Core Notes

3.1 Problem

Prompt-level safety is not enough for agents with file access, shell access, credentials, private data, and long-running memory.

对于拥有文件访问、shell 访问、凭证、私有数据和长期记忆的智能体，仅靠提示词安全是不够的。

  • Framework choice does not solve runtime security. 框架选择不能解决运行时安全。

  • Model serving does not solve workflow governance. 模型服务不能解决工作流治理。

  • Private data agents require source traceability. 私有数据智能体需要来源可追溯。

3.2 Mechanism

The stack separates concerns: NIM for inference, NeMo Agent Toolkit for workflow and observability, AI-Q for enterprise retrieval patterns, and OpenShell for sandbox policy.

这个栈分离关注点：NIM 负责推理，NeMo Agent Toolkit 负责工作流和可观测性，AI-Q 负责企业检索模式，OpenShell 负责沙箱策略。

  • Use NIM as repeatable model-serving unit. 把 NIM 作为可复用模型服务单元。

  • Use toolkit evaluation and telemetry to improve agents. 用工具包评估和遥测改进智能体。

  • Use runtime policy below the agent. 在智能体下层使用运行时策略。

3.3 Evidence

NVIDIA materials describe Agent Toolkit support for frameworks, MCP, A2A, profiling, evaluation, observability, NIM integration, and secure runtimes through OpenShell.

NVIDIA 材料描述了 Agent Toolkit 对框架、MCP、A2A、画像、评估、可观测性、NIM 集成和 OpenShell 安全运行时的支持。

  • AI-Q connects enterprise data, retrieval, reasoning, and report generation. AI-Q 连接企业数据、检索、推理和报告生成。

  • OpenShell enforces policy outside the model. OpenShell 在模型外部执行策略。

  • NemoClaw is a reference stack, not the runtime itself. NemoClaw 是参考栈，不是运行时本身。

3.4 Boundary

NemoClaw and related always-on agent patterns require maturity validation before enterprise rollout.

NemoClaw 和相关常驻智能体模式在企业推广前需要成熟度验证。

  • Do not expose raw credentials to agents. 不要把原始凭证暴露给智能体。

  • Do not skip evaluation datasets. 不要跳过评估数据集。

  • Do not confuse framework interoperability with security. 不要把框架互操作性误认为安全。

4. Concept Map

Use wikilinks to connect this note into the broader Quartz graph.

使用双向链接把这篇笔记接入更大的 Quartz 知识网络。

flowchart LR
  A["Enterprise Agent Need"] --> B["NIM"]
  A --> C["NeMo Agent Toolkit"]
  A --> D["AI-Q Blueprint"]
  A --> E["OpenShell"]
  B --> F["Model Serving"]
  C --> G["Workflow and Evaluation"]
  D --> H["Grounded Research"]
  E --> I["Runtime Policy"]
  F --> J["Production Agent Operations"]
  G --> J
  H --> J
  I --> J

Diagram labels stay in English for rendering consistency and easier reuse across published pages.

图中的标签保持英文，便于 Quartz 渲染后跨页面复用，也方便技术读者快速识别。

5. My Take

The most important pattern is separation of concerns. Enterprise agents need model serving, workflow orchestration, data retrieval, runtime security, and observability as independent but integrated layers.

最重要的模式是关注点分离。企业智能体需要把模型服务、工作流编排、数据检索、运行时安全和可观测性作为独立但集成的层来处理。

  • What changed my thinking: Runtime control belongs below the agent. 改变我理解的地方：运行时控制应该在智能体下层。

  • What I may do next: Design a private manufacturing research agent using separate inference, retrieval, and policy layers. 下一步可能行动：用独立推理、检索和策略层设计私有制造研究智能体。

  • What still needs verification: Product maturity, licensing, deployment path, and integration burden. 仍需要验证的内容：产品成熟度、许可、部署路径和集成负担。

Reuse Path

Convert this note into an enterprise agent platform reference architecture.

References

Graph View

Backlinks