Google

Vendor profile — Google. Reviewed 2026-10-18 from vendor documentation only; no hands-on testing.

厂商档案 — Google。于 2026-10-18 依据厂商文档复核；没有实际操作测试。

Profile概况

Organization机构
Google (Google Cloud)
Category类别
Hyperscaler that is also a frontier model lab: one cloud platform for models, agents and classical ML (interpretation).既是超大规模云、也是前沿模型实验室：用一个云平台覆盖模型、智能体和传统 ML（解读）。
Primary buyer主要买家
Developers and ML teams (build, evaluate, deploy) and enterprise security administrators (Agent Gateway, Agent Identity, Model Armor) [1][6]. Observation from the audiences the pages address.开发者与 ML 团队（构建、评测、部署），以及企业安全管理员（Agent Gateway、Agent Identity、Model Armor）[1][6]。依据页面面向的读者所作的观察。
Relevant products相关产品
Gemini Enterprise Agent Platform, the renamed Vertex AI [1][2]: Agent Development Kit, Agent Studio, Agent Runtime, Agent Registry, Agent Identity, Agent Gateway, Model Armor [1][6][7]; Gen AI and agent evaluation [3][4]; Observability [5]; Model Registry, Pipelines, Model Monitoring [9][10][8].Gemini Enterprise Agent Platform（即改名后的 Vertex AI）[1][2]：Agent Development Kit、Agent Studio、Agent Runtime、Agent Registry、Agent Identity、Agent Gateway、Model Armor [1][6][7]；Gen AI 与智能体评测 [3][4]；可观测 [5]；Model Registry、Pipelines、Model Monitoring [9][10][8]。
Last reviewed最后复核
2026-10-18

MDLC coverageMDLC 阶段覆盖

Coverage by stage按阶段的覆盖情况 — ● full完整 · ◐ partial部分 · ○ none无 · ? not yet researched尚未研究
Stage阶段Coverage覆盖Product / feature产品 / 功能Source来源
Require需求?No requirements or intake feature found in the pages read. Not researched in depth.所读页面中没有发现需求/立项功能。未深入研究。—
Plan规划◐Model Garden gives access to Google, third-party and open models (the overview cites more than 200 foundation models) [1]. The Gen AI evaluation service supports head-to-head comparison of Google and third-party models on your own data to pick the best fit [3]. No tool-set or autonomy-level planning feature found.Model Garden 提供 Google、第三方与开源模型（概览称超过 200 个基础模型）[1]。Gen AI 评测服务支持在你自己的数据上对 Google 与第三方模型做对比，以选出最合适的 [3]。没有发现工具集或自主程度的规划功能。[1][3]
Build构建●Agents: Agent Development Kit (model-agnostic), Agent Studio (low-code canvas), Agent Garden templates, a Managed Agents API in a managed sandbox, RAG Engine, Vector Search [1]. Classical ML: training pipelines, Pipelines on Kubeflow or TFX with ML Metadata lineage, Colab Enterprise notebooks [1][10]. The agent-evaluation page mentions Gemini CLI as an AI coding assistant that can run evaluation skills [4]; a coding-agent product itself was not researched.智能体：Agent Development Kit（模型无关）、Agent Studio（低代码画布）、Agent Garden 模板、受管沙箱中的 Managed Agents API、RAG Engine、Vector Search [1]。传统 ML：训练管线、基于 Kubeflow 或 TFX 的 Pipelines 与 ML Metadata 血缘、Colab Enterprise 笔记本 [1][10]。智能体评测页提到 Gemini CLI 作为可运行评测技能的 AI 编码助手 [4]；编码智能体产品本身未研究。[1][10][4]
Validate验证●Gen AI evaluation service: adaptive rubrics (recommended), static rubrics, computation-based metrics, custom Python functions; datasets from uploads, sampled production logs or synthetic data [3]. Agent evaluation: scenario generation, multi-turn user simulation, environment simulation (injected tool errors), multi-turn autoraters, prompt optimisation, evaluation of production traces [4]. Platform also lists AI threat and vulnerability scanning for agentic systems [1].Gen AI 评测服务：自适应评分标准（推荐）、静态评分标准、基于计算的指标、自定义 Python 函数；数据集可来自上传、抽样生产日志或合成数据 [3]。智能体评测：场景生成、多轮用户模拟、环境模拟（注入工具错误）、多轮自动评分器、提示词优化、对生产轨迹的评估 [4]。平台还列有面向智能体系统的 AI 威胁与漏洞扫描 [1]。[3][4][1]
Approve审批?Agent Registry is described as a catalogue of approved agents, tools and MCP servers that Agent Gateway checks before allowing connections [6], but that is destination approval at run time, not a release sign-off. A “Share agent” page exists in navigation (title only). Not marked.Agent Registry 被描述为已批准智能体、工具与 MCP 服务器的目录，Agent Gateway 在放行连接前查询它 [6]，但这是运行时的目的地批准，不是发布签批。导航中有“Share agent”页面（只见标题）。不打分。[6]
Deploy部署●Agent Runtime: managed, stateful runtime with sub-second cold starts and long-running agents; five deployment routes (execution object, source files, Dockerfile, Artifact Registry image, Git repository); supports ADK, A2A, LangChain, LangGraph, AG2, LlamaIndex and custom agents [1][7]. Application Design Center templates include an “Agent Platform with governance” variant [7]. Classical ML: assign a model version to an endpoint from Model Registry by alias, or run batch inference [9].Agent Runtime：受管、有状态的运行时，亚秒级冷启动，支持长时间运行的智能体；五种部署路径（执行对象、源文件、Dockerfile、Artifact Registry 镜像、Git 仓库）；支持 ADK、A2A、LangChain、LangGraph、AG2、LlamaIndex 与自定义智能体 [1][7]。Application Design Center 模板含“带治理的 Agent Platform”版本 [7]。传统 ML：在 Model Registry 中用别名把模型版本指派到端点，或运行批量推理 [9]。[1][7][9]
Monitor监控●Agent observability on OpenTelemetry telemetry in Google Cloud Observability: usage, token, latency percentile and error dashboards; online monitors for response quality, safety, hallucination and tool-use quality; per-model and per-tool views; traces and topology [5]. Model Monitoring for tabular models: v1 is generally available on endpoints, v2 is preview; drift tracking with threshold alerts [8].基于 Google Cloud Observability 中 OpenTelemetry 遥测的智能体可观测：用量、token、延迟分位数与错误率仪表板；针对回答质量、安全、幻觉与工具使用质量的在线监控；按模型和按工具的视图；追踪与拓扑 [5]。表格型模型的 Model Monitoring：v1 在端点上正式可用，v2 为预览；漂移跟踪与阈值告警 [8]。[5][8]
Governance治理●Agent Identity (a SPIFFE ID per agent), Agent Registry, Agent Gateway as the policy enforcement point with default-deny IAM policies, Model Armor filters for prompt injection and data leaks, and plain-language Semantic Governance policies [6][1]. Model Registry models are searchable across projects in Knowledge Catalog [9]. Pipelines track artifact lineage with ML Metadata [10]. The audit section of the observability page was cut off in the read, so audit claims are not made here.Agent Identity（每个智能体一个 SPIFFE ID）、Agent Registry、作为策略执行点的 Agent Gateway（默认拒绝的 IAM 策略）、拦截提示词注入与数据泄露的 Model Armor 过滤器，以及用自然语言编写的语义治理策略 [6][1]。Model Registry 中的模型可在 Knowledge Catalog 中跨项目搜索 [9]。Pipelines 用 ML Metadata 跟踪制品血缘 [10]。可观测页的审计部分在读取时被截断，因此这里不对审计作论断。[6][1][9][10]

Observations观察

Google renamed Vertex AI to Gemini Enterprise Agent Platform. Examples from its rename table: Vertex AI Agent Engine is now Agent Runtime, Vertex AI Studio is Agent Studio, Vertex AI Model Garden is Gemini Enterprise Agent Platform Model Garden. [2] Older notes and sources on this site (for example S-0030, S-0031) use the old names.

Google 把 Vertex AI 改名为 Gemini Enterprise Agent Platform。改名表中的例子：Vertex AI Agent Engine 现为 Agent Runtime，Vertex AI Studio 现为 Agent Studio，Vertex AI Model Garden 现为 Gemini Enterprise Agent Platform Model Garden。[2] 本站较早的笔记与来源（如 S-0030、S-0031）仍使用旧名称。

Agent Gateway blocks all connections by default unless an IAM policy grants access, and attaches Model Armor to scan prompts and tool responses [6].

Agent Gateway 默认阻断所有连接，除非有 IAM 策略授权，并挂接 Model Armor 扫描提示词与工具响应 [6]。

Model Monitoring covers tabular models only; v2 is preview, v1 is generally available and recommended for production endpoints [8]. Agent evaluation and observability pages carry no preview label in the text read, but status was not checked on a separate release-notes page.

Model Monitoring 只覆盖表格型模型；v2 为预览，v1 正式可用，生产端点推荐 v1 [8]。所读文本中智能体评测与可观测页没有预览标注，但没有在发布说明页核对状态。

Method and limits: vendor documentation only, no hands-on testing. Pages were read with the built-in browser (main article text), but several were cut at a character limit (Agent Gateway, Observability, Pipelines, Agent evaluation), so the later parts were not read. Marks mean a first-party feature is documented, not that it is good or widely used.

方法与局限：仅读厂商文档，没有实际操作。页面通过内置浏览器读取正文，但有几页在字数上限处被截断（Agent Gateway、可观测、Pipelines、智能体评测），后半部分未读。标记只表示文档中存在厂商自有功能，不代表质量或使用规模。

Interpretation解读

Google documents an explicit agent-governance chain: identity per agent, a registry, a gateway enforcing default-deny policy, and content filters, all attached to the runtime. That supports the view that governance converges in a control plane (H4 on the thesis page), but note that here it is a runtime enforcement point (policy at the network edge), which is a different shape from an approval-and-evidence register.

Google 的文档描述了一条明确的智能体治理链条：每个智能体有身份、有注册表、有按默认拒绝执行策略的网关，以及内容过滤，全部挂在运行时上。这支持“治理在控制平面汇聚”的看法（论点页 H4），但要注意，这里是运行时的执行点（网络边缘上的策略），与“审批加证据的登记簿”形态不同。

Agent evaluation here is built around synthetic scenarios, user simulation and automated raters, with an optimisation loop that proposes prompt changes. That is closer to the “workflow structure vs model choice” question (H3 on the thesis page) than a plain test suite, but nothing in the pages measures it.

这里的智能体评测围绕合成场景、用户模拟和自动评分器，并带有提出提示词修改的优化循环。比起普通测试集，它更贴近“工作流结构 vs 模型选择”的问题（论点页 H3），但页面中没有任何对其效果的度量。

Strategic position战略位置

FACT. The platform is described as organised around Build, Scale, Govern and Optimize [1].

FACT。平台被描述为围绕 Build、Scale、Govern、Optimize 四个支柱组织 [1]。

INTERPRETATION. Strengths: one platform spanning classical ML (registry, pipelines, monitoring) and agents (runtime, evaluation, gateway); explicit agent identity and gateway design; own models plus third-party models in the same catalogue. Gaps: no requirements stage or release-approval workflow found; Model Monitoring v2 and parts of the tabular stack are preview; a large rename may leave docs and customer knowledge split between old and new names.

INTERPRETATION。优势：同一平台覆盖传统 ML（注册表、管线、监控）与智能体（运行时、评测、网关）；明确的智能体身份与网关设计；自有模型与第三方模型在同一目录。缺口：没有发现需求阶段或发布审批流程；Model Monitoring v2 与部分表格型栈仍为预览；大规模改名可能让文档和客户认知在新旧名称之间割裂。

HYPOTHESIS. Google is positioning the gateway-plus-identity layer as the place where enterprise agent policy is enforced regardless of where the agent was built (the docs say the gateway is framework-agnostic [6]). Evidence that would test this: customer examples of non-Google agents governed through it.

HYPOTHESIS。Google 把“网关加身份”这一层定位为不论智能体在哪里构建、企业策略都在此执行的地方（文档称网关与框架无关 [6]）。检验证据：非 Google 智能体经由它受管的客户案例。

Open questions待解问题

What does the audit-logging section of the Observability page say about auditing agent actions (the text was cut off)? This bears on Q2 (which systems own agent action logs).

可观测页关于智能体动作审计的部分写了什么（读取时被截断）？这关系到 Q2（谁来承载智能体动作日志）。

Is there a release-approval or launch-review feature (Agent Registry states, “Share agent”) beyond runtime allow-lists?

除了运行时的允许清单外，是否有发布审批或上线评审功能（Agent Registry 状态、“Share agent”）？

Which agent-evaluation and observability pieces are GA versus preview? Check release notes.

智能体评测与可观测的哪些部分是正式可用、哪些是预览？需查发布说明。

Require is “?”, and a coding-agent product (Gemini CLI, Jules) was not researched for the Build row.

Require 仍为“?”；Build 行里编码智能体产品（Gemini CLI、Jules）未研究。

References参考来源

  1. Google Cloud, “Agent Platform overview,” Gemini Enterprise Agent Platform docs (last updated 2026-10-16). https://docs.cloud.google.com/gemini-enterprise-agent-platform/overview vendor documentation · S-0088
  2. Google Cloud, “Gemini Enterprise Agent Platform name changes” (Vertex AI renames). https://docs.cloud.google.com/gemini-enterprise-agent-platform/vertex-ai-name-changes vendor documentation · S-0089
  3. Google Cloud, “Gen AI evaluation service overview.” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/evaluation-overview vendor documentation · S-0090
  4. Google Cloud, “Agent evaluation.” https://docs.cloud.google.com/gemini-enterprise-agent-platform/optimize/evaluation/agent-evaluation vendor documentation · S-0091
  5. Google Cloud, “Observability overview.” https://docs.cloud.google.com/gemini-enterprise-agent-platform/optimize/observability/overview vendor documentation · S-0092
  6. Google Cloud, “Agent Gateway overview.” https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/gateways/agent-gateway-overview vendor documentation · S-0093
  7. Google Cloud, “Deploy an agent” (Agent Runtime; last updated 2026-10-15). https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/deploy-an-agent vendor documentation · S-0094
  8. Google Cloud, “Introduction to Model Monitoring.” https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/model-monitoring/overview vendor documentation · S-0095
  9. Google Cloud, “Introduction to Model Registry” (last updated 2026-10-15). https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/model-registry/introduction vendor documentation · S-0096
  10. Google Cloud, “Introduction to Gemini Enterprise Agent Platform Pipelines.” https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/pipelines/introduction vendor documentation · S-0097