Microsoft

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

厂商档案 — Microsoft。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
Microsoft
Category类别
Hyperscaler and platform vendor: cloud AI platform (Microsoft Foundry, Azure Machine Learning), enterprise identity and compliance stack (Entra, Purview, Defender), developer tools (GitHub Copilot).超大规模云与平台厂商：云 AI 平台（Microsoft Foundry、Azure Machine Learning）、企业身份与合规套件（Entra、Purview、Defender）、开发者工具（GitHub Copilot）。
Primary buyer主要买家
Developers and AI platform teams (Foundry), IT and security administrators (Agent 365, Control Plane, Purview) [3][8]. Observation from the stated audience of the Learn pages.开发者与 AI 平台团队（Foundry）；IT 与安全管理员（Agent 365、Control Plane、Purview）[3][8]。依据 Learn 页面写明的目标读者所作的观察。
Relevant products相关产品
Microsoft Foundry: Agent Service, model catalog, evaluation, observability, Control Plane [1][2][3]; Microsoft Agent 365 and Entra Agent ID [4][8]; Azure Machine Learning MLOps and Responsible AI dashboard [6][9]; AI Red Teaming Agent [10]; Purview for AI [12]; GitHub Copilot cloud agent [13].Microsoft Foundry：Agent Service、模型目录、评测、可观测、Control Plane [1][2][3]；Microsoft Agent 365 与 Entra Agent ID [4][8]；Azure Machine Learning 的 MLOps 与 Responsible AI 仪表板 [6][9]；AI Red Teaming Agent [10]；Purview for AI [12]；GitHub Copilot cloud agent [13]。
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 (no review of Microsoft’s Responsible AI Standard templates), so “?”.所读页面中没有发现需求/立项功能。未深入研究（未查看 Microsoft 负责任 AI 标准的模板），因此为“?”。—
Plan规划◐Foundry Models catalog with more than 10,000 models from Microsoft, OpenAI, Anthropic, Meta and others [1]. Base-model selection by comparing quality, task performance and safety, with the Foundry benchmark on public datasets or your own data [2]. No tool-set or autonomy-level planning feature found.Foundry 模型目录有 10,000 多个来自 Microsoft、OpenAI、Anthropic、Meta 等的模型 [1]。通过比较质量、任务表现与安全性来选基础模型，Foundry 基准可用公开数据集或自有数据 [2]。没有发现工具集或自主程度的规划功能。[1][2]
Build构建●Foundry Agent Service: prompt agents, voice agents and containerised hosted agents (Microsoft Agent Framework, LangGraph, Semantic Kernel), immutable agent versions, Toolbox [1][5]. Supervised fine-tuning with portal, SDK, REST and azd, training metrics and checkpoints [11]. Azure ML pipelines, environments and model registry for classical ML [6]. GitHub Copilot cloud agent: background coding agent, one pull request per task, sessions up to 59 minutes [13]. Covers the three system types in the repo’s working model (classical ML, LLM app, agent) plus coding agents; some parts are preview [1].Foundry Agent Service：提示词智能体、语音智能体与容器化托管智能体（Microsoft Agent Framework、LangGraph、Semantic Kernel）、不可变的智能体版本、Toolbox [1][5]。监督微调支持门户、SDK、REST 与 azd，含训练指标与检查点 [11]。Azure ML 管线、环境与模型注册表（传统 ML）[6]。GitHub Copilot cloud agent：后台编码智能体，每个任务一个 PR，单次会话最长 59 分钟 [13]。覆盖本仓库工作模型中的三类系统（传统 ML、LLM 应用、智能体）及编码智能体；部分为预览 [1]。[1][5][11][6][13]
Validate验证●Foundry built-in evaluators (coherence, fluency, groundedness, relevance, safety, tool call accuracy, task completion) plus custom evaluators; pre-production evaluation and CI/CD quality gates [2]. AI Red Teaming Agent built on PyRIT, reporting attack success rate and a scorecard to decide deployment readiness [10]. Azure ML Responsible AI dashboard for tabular classification and regression models: error analysis, fairness, interpretability, counterfactuals, PDF scorecard [9]. The lifecycle page labels agent tracing and evaluation “preview” [5].Foundry 内置评估器（连贯性、流畅性、有据性、相关性、安全、工具调用准确率、任务完成度）及自定义评估器；上线前评测与 CI/CD 质量门禁 [2]。基于 PyRIT 的 AI Red Teaming Agent，输出攻击成功率与评分卡，用于判断是否可部署 [10]。Azure ML 的 Responsible AI 仪表板，适用于表格型分类与回归模型：错误分析、公平性、可解释性、反事实、PDF 评分卡 [9]。生命周期页面把智能体追踪与评测标为“预览” [5]。[2][10][9][5]
Approve审批◐A Foundry hosted agent published as an Agent 365 “autopilot” blueprint appears in the registry after administrator approval [4]. The Responsible AI scorecard is described as something to share with compliance stakeholders “to get their approval” [9]. Control Plane applies versioned policies with assignment tracking [3]. No general model-release sign-off workflow found in the pages read.作为 Agent 365“autopilot”蓝图发布的 Foundry 托管智能体，经管理员批准后才出现在注册表 [4]。Responsible AI 评分卡被描述为可交给合规相关方以“获得批准”的材料 [9]。Control Plane 应用带版本的策略并跟踪分配 [3]。所读页面中没有发现通用的模型发布签批流程。[4][9][3]
Deploy部署●Foundry deployment types: Global, Data Zone and geography Standard; Provisioned throughput; Batch at 50% off; a Developer type for fine-tuned model evaluation; Azure Policy can restrict types [7]. Agents publish as a stable “agent application” endpoint; hosted agents get a managed endpoint [5][1]. Azure ML managed online and batch endpoints with traffic-split A/B rollout [6].Foundry 部署类型：Global、Data Zone 与按地理区域的 Standard；预置吞吐；五折的 Batch；用于评估微调模型的 Developer 类型；可用 Azure Policy 限制类型 [7]。智能体可发布为稳定的“agent application”端点；托管智能体获得受管端点 [5][1]。Azure ML 受管在线与批处理端点，支持流量切分的 A/B 发布 [6]。[7][5][1][6]
Monitor监控●Foundry dashboards on Azure Monitor Application Insights: token use, latency, error rates, quality scores, alerts; continuous evaluation of sampled production traffic; scheduled evaluation and scheduled red teaming; OpenTelemetry tracing [2]. Control Plane fleet view with cost and rate-limit anomalies [3]. Azure ML publishes drift and lifecycle events to Event Grid [6]. Dashboards are labelled preview on the overview page [1].Foundry 基于 Azure Monitor Application Insights 的仪表板：token 用量、延迟、错误率、质量分、告警；对抽样的生产流量持续评估；定时评测与定时红队；OpenTelemetry 追踪 [2]。Control Plane 的舰队视图含成本与限流异常 [3]。Azure ML 把漂移与生命周期事件发布到 Event Grid [6]。概览页把仪表板标为预览 [1]。[2][3][6][1]
Governance治理●Foundry Control Plane: one inventory of agents, models and tools across projects, guardrail policies, a Compliance pane integrating Azure Policy, Defender and Purview, bulk remediation [3]. Microsoft Agent 365: registry, Entra Agent ID identities, access control, security; generally available 2026-05-01 for the Commercial segment, per user [8][4]. Purview audit, DLP, retention, eDiscovery and Compliance Manager for Copilot and other AI apps, including Foundry and ChatGPT Enterprise [12]. Licensing gates apply (see Observations).Foundry Control Plane：跨项目统一登记智能体、模型与工具，护栏策略，整合 Azure Policy、Defender 与 Purview 的合规面板，批量整改 [3]。Microsoft Agent 365：注册表、Entra Agent ID 身份、访问控制、安全；2026-05-01 起面向商业版按用户计费正式可用 [8][4]。Purview 的审计、DLP、保留、电子取证与 Compliance Manager 覆盖 Copilot 及其他 AI 应用，包括 Foundry 与 ChatGPT Enterprise [12]。受许可条件限制（见观察）。[3][8][4][12]

Observations观察

Agent 365 reached general availability on 2026-05-01 for the Commercial segment, licensed per user; Microsoft says it works best with Microsoft E5 and needs at least one qualifying Agent 365 license [8]. The Foundry integration page adds that, before Foundry sends agent activity data to Agent 365, the tenant needs at least one Microsoft 365 Copilot license and enrolment in the Frontier preview program, and a global admin must accept terms [4].

Agent 365 于 2026-05-01 面向商业版正式可用，按用户授权；Microsoft 称配合 Microsoft E5 效果最好，且至少需要一个合格的 Agent 365 许可证 [8]。Foundry 集成页补充：Foundry 把智能体活动数据送入 Agent 365 之前，租户至少需要一个 Microsoft 365 Copilot 许可证并加入 Frontier 预览计划，全局管理员还须接受条款 [4]。

The two Learn pages frame the prerequisites differently (a generally available product versus a preview-program enrolment), so the practical entry cost for a Foundry customer is unclear from documentation alone.

两个 Learn 页面对前置条件的说法不同（正式可用产品 vs. 需加入预览计划），仅凭文档无法确定 Foundry 客户实际的准入成本。

Foundry’s own pages mark several lifecycle capabilities as preview: dashboards on the overview page [1], and agent tracing and evaluation on the lifecycle page [5]. The Control Plane page notes that preview items have no SLA and are not recommended for production workloads [3].

Foundry 自己的页面把若干生命周期能力标为预览：概览页的仪表板 [1]，生命周期页的智能体追踪与评测 [5]。Control Plane 页说明预览项没有 SLA，不建议用于生产 [3]。

Foundry Agent Service names the Responses API (Agents v2) as the successor of the Assistants API, and tracing supports the OpenAI Agents SDK among other frameworks [1][2]. Control Plane can register agents built elsewhere [3], though the AI Red Teaming Agent currently supports only Foundry hosted prompt and container agents [10].

Foundry Agent Service 以 Responses API（Agents v2）取代 Assistants API，追踪支持包括 OpenAI Agents SDK 在内的多种框架 [1][2]。Control Plane 可登记其他平台构建的智能体 [3]，但 AI Red Teaming Agent 目前只支持 Foundry 托管的提示词与容器智能体 [10]。

Method and limits: vendor documentation only, no hands-on testing. The Microsoft Learn pages were returned as full text by the fetch tool (including their front matter and ms.date); the GitHub page was read through a summariser. Coverage marks reflect that a first-party feature is documented, not its quality or how many customers use it.

方法与局限：仅读厂商文档，没有实际操作。Microsoft Learn 页面由抓取工具返回全文（含 front matter 与 ms.date）；GitHub 页面通过摘要器读取。覆盖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

Microsoft is the only one of the two reviewed vendors whose documents describe a path through all of Plan, Build, Validate, Deploy, Monitor and Governance for classical ML, LLM apps and agents, plus an identity-based registry of agents (Agent 365 with Entra Agent ID). This lines up with the thesis claim that the control plane (registry, approval, evidence) is where tooling converges (H4 on the thesis page); it is support from vendor self-description, not independent evidence.

在已复核的两家厂商中，只有 Microsoft 的文档描述了覆盖传统 ML、LLM 应用与智能体的 Plan、Build、Validate、Deploy、Monitor、Governance 全链路，还有基于身份的智能体注册表（Agent 365 加 Entra Agent ID）。这与论点“控制平面（注册、审批、证据）是工具汇聚之处”（论点页 H4）相吻合；但这只是厂商自述的支持，不是独立证据。

Microsoft’s agent versions are immutable and requests can be pinned to a version [5]. That is a small data point for the idea that agent configuration is versioned like code (H2 on the thesis page).

Microsoft 的智能体版本不可变，请求可固定到某个版本 [5]。这是“智能体配置像代码一样做版本管理”（论点页 H2）的一个小证据。

Strategic position战略位置

FACT. Control Plane and Agent 365 integrate with Entra, Defender and Purview [3][8]; Purview lists ChatGPT Enterprise and Claude Enterprise as supported enterprise AI apps [12].

FACT。Control Plane 与 Agent 365 与 Entra、Defender、Purview 集成 [3][8]；Purview 把 ChatGPT Enterprise 与 Claude Enterprise 列为受支持的企业 AI 应用 [12]。

INTERPRETATION. Strengths: breadth across the lifecycle, an existing enterprise identity and compliance estate that agents can plug into, and a neutral stance on models (catalog includes OpenAI, Anthropic, Meta). Gaps: no requirements stage found; several lifecycle pieces still in preview; the strongest governance features depend on licences and tenant set-up; red teaming is limited to Foundry-hosted agents.

INTERPRETATION。优势：生命周期覆盖面广；智能体可以接入已有的企业身份与合规体系；对模型保持中立（目录含 OpenAI、Anthropic、Meta）。缺口：没有发现需求阶段；多个环节仍在预览；最强的治理功能取决于许可证与租户配置；红队测试仅限 Foundry 托管智能体。

HYPOTHESIS. Buyers already on Microsoft 365 E5 will treat Agent 365 and Purview as the default register and audit layer even for agents built on other vendors’ models. Evidence that would test this: customer reports of agents from non-Microsoft stacks actually registered and governed there.

HYPOTHESIS。已使用 Microsoft 365 E5 的买家，即使智能体基于其他厂商的模型，也会把 Agent 365 与 Purview 作为默认的登记与审计层。检验证据：客户报告中非 Microsoft 技术栈的智能体确实在其中登记并受管。

Open questions待解问题

Which Foundry observability and evaluation pieces are generally available today? The pages disagree on preview labels; check the general-availability overview page the docs link to.

目前 Foundry 的可观测与评测哪些已正式可用？各页的预览标注不一致；需核对文档所链接的“general availability overview”页。

What does an Agent 365 “approval pipeline” look like in practice (owner, sponsor, approvers)? Only the autopilot publishing approval was confirmed; the Learn training module that mentions it was not read.

Agent 365 的“审批流水线”实际是什么样（所有者、担保人、审批人）？目前只确认了 autopilot 发布需管理员批准；提到它的 Learn 培训模块尚未阅读。

How well do Control Plane and Agent 365 cover agents built outside Foundry, for example on OpenAI or Anthropic stacks? (Registration is documented; depth of evaluation and policy enforcement is not.)

Control Plane 与 Agent 365 对 Foundry 之外构建的智能体（如基于 OpenAI 或 Anthropic 的栈）覆盖到什么程度？（登记有文档；评测与策略执行的深度没有。）

Require is “?”, and Microsoft’s Responsible AI Standard impact-assessment templates were not reviewed; do they count as a Require/Approve artefact?

Require 仍为“?”，Microsoft 负责任 AI 标准的影响评估模板未复核；它们能否算作 Require/Approve 的产物？

References参考来源

  1. Microsoft Learn, “What is Microsoft Foundry?” (ms.date 2026-08-13). https://learn.microsoft.com/en-us/azure/foundry/what-is-foundry vendor documentation · S-0075
  2. Microsoft Learn, “Observability in Generative AI” (Foundry; ms.date 2026-10-15). https://learn.microsoft.com/en-us/azure/foundry/concepts/observability vendor documentation · S-0076
  3. Microsoft Learn, “What is Microsoft Foundry Control Plane?” (ms.date 2026-05-06). https://learn.microsoft.com/en-us/azure/foundry/control-plane/overview vendor documentation · S-0077
  4. Microsoft Learn, “Microsoft Agent 365 integration with Foundry” (ms.date 2026-08-26). https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/agent-365-integration vendor documentation · S-0078
  5. Microsoft Learn, “Agent development lifecycle in Microsoft Foundry” (ms.date 2026-08-21). https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/development-lifecycle vendor documentation · S-0079
  6. Microsoft Learn, “MLOps machine learning model management,” Azure Machine Learning (ms.date 2025-10-06). https://learn.microsoft.com/en-us/azure/machine-learning/concept-model-management-and-deployment vendor documentation · S-0080
  7. Microsoft Learn, “Understanding deployment types in Microsoft Foundry Models” (ms.date 2026-08-06). https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/deployment-types vendor documentation · S-0081
  8. Microsoft Learn, “Microsoft Agent 365 overview” (ms.date 2026-08-19). https://learn.microsoft.com/en-us/microsoft-agent-365/overview vendor documentation · S-0082
  9. Microsoft Learn, “Responsible AI dashboard,” Azure Machine Learning (ms.date 2025-10-13). https://learn.microsoft.com/en-us/azure/machine-learning/concept-responsible-ai-dashboard vendor documentation · S-0083
  10. Microsoft Learn, “AI Red Teaming Agent,” Foundry (ms.date 2026-08-19). https://learn.microsoft.com/en-us/azure/foundry/concepts/ai-red-teaming-agent vendor documentation · S-0084
  11. Microsoft Learn, “Fine-tune a model with supervised fine-tuning in Microsoft Foundry” (ms.date 2026-10-05). https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/fine-tuning vendor documentation · S-0085
  12. Microsoft Learn, “Microsoft Purview data security and compliance protections for Microsoft 365 Copilot and other generative AI apps” (ms.date 2026-05-27). https://learn.microsoft.com/en-us/purview/ai-microsoft-purview vendor documentation · S-0086
  13. GitHub Docs, “About GitHub Copilot cloud agent” (GitHub is a Microsoft company; read through a summariser). https://docs.github.com/en/copilot/concepts/agents/coding-agent/about-coding-agent vendor documentation (via summariser) · S-0087