Anthropic

Vendor profile — Anthropic. Reviewed 2026-10-18 from vendor documentation only; no hands-on testing. Disclosure: written with Claude, an Anthropic model.

厂商档案 — Anthropic。于 2026-10-18 依据厂商文档复核；没有实际操作测试。披露：由 Anthropic 的模型 Claude 协助撰写。

Profile概况

Organization机构
Anthropic
Category类别
Frontier lab selling models through an API, agent infrastructure (Managed Agents, beta) and a coding agent (Claude Code), also available through AWS, Google Cloud and Microsoft Foundry [11].前沿模型实验室：通过 API 销售模型、提供智能体基础设施（Managed Agents，测试版）和编码智能体（Claude Code），也可通过 AWS、Google Cloud 与 Microsoft Foundry 使用 [11]。
Primary buyer主要买家
Developers and engineering teams (API, Claude Code), and organisation owners and admins (analytics, Compliance API, managed settings) [5][7][4][9]. Observation from who the pages address.开发者与工程团队（API、Claude Code），以及组织所有者与管理员（分析、合规 API、托管设置）[5][7][4][9]。依据页面面向的读者所作的观察。
Relevant products相关产品
Claude Managed Agents (beta) [1]; Claude Code, Code Review, analytics, managed settings [5][8][7][9]; Compliance API [4]; Usage and Cost reporting [12].Claude Managed Agents（测试版）[1]；Claude Code、Code Review、分析、托管设置 [5][8][7][9]；合规 API [4]；用量与成本报表 [12]。
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?Only guidance, no tool: the evaluation page advises that success criteria be specific, measurable, achievable and relevant [10]. Not marked.只有指导、没有工具：评测页建议成功标准应具体、可衡量、可达成且相关 [10]。不打分。[10]
Plan规划?No planning feature found. The model-status page gives retirement dates and replacements, with at least 60 days’ notice for publicly released models [11], which informs migration planning but is not a planning tool. Not marked.没有发现规划功能。模型状态页给出退役日期与替代模型，公开发布的模型至少提前 60 天通知 [11]，这有助于迁移规划，但不是规划工具。不打分。[11]
Build构建●Claude Code: an agentic coding tool for terminal, IDE, desktop app and web, with CLAUDE.md instructions, skills, hooks, subagents, MCP and an Agent SDK for custom agents [5]. Claude Managed Agents (beta): a managed agent harness (agent, environment, session, events) with built-in bash, file, web and MCP tools [1]. Interpretation: strongest in coding agents; no classical-ML training or pipeline product was found, and none is claimed.Claude Code：面向终端、IDE、桌面应用与网页的智能体编码工具，带 CLAUDE.md 指令、skills、hooks、subagents、MCP 与用于自建智能体的 Agent SDK [5]。Claude Managed Agents（测试版）：受管的智能体运行框架（agent、environment、session、events），内置 bash、文件、网页与 MCP 工具 [1]。解读：强项在编码智能体；没有发现传统 ML 训练或管线产品，也不作此声称。[5][1]
Validate验证◐Guidance for building evaluations (task-specific cases, code-graded, embedding-based and LLM-graded examples) [10], but the documentation index lists no evaluation-tool page [13]. Code Review: several agents read a GitHub pull request, a verification step filters false positives, findings come as inline comments; Team and Enterprise plans [8]. It reviews code; it does not evaluate an agent or model.构建评测的指导（任务相关用例、代码评分、基于嵌入与 LLM 评分的示例）[10]，但文档索引中没有评测工具页面 [13]。Code Review：多个智能体阅读 GitHub 拉取请求，验证步骤过滤误报，结果以行内评论给出；面向 Team 与 Enterprise 计划 [8]。它评审的是代码，不是评估智能体或模型。[10][13][8]
Approve审批◐Managed Agents permission policies: always-allow, or always-ask so the session pauses for your approval before a tool runs; custom tools are controlled by your own application [2]. Code Review never approves or blocks a pull request, so the merge decision stays with the team [8]. Claude Code also lists permission-mode and permissions pages [14] (titles only). Human approval of agent actions, not a release sign-off workflow.Managed Agents 权限策略：始终允许，或“始终询问”，使会话暂停并等你批准后才执行工具；自定义工具由你自己的应用控制 [2]。Code Review 从不批准或阻止拉取请求，合并决定仍由团队做出 [8]。Claude Code 还列有权限模式与权限页面 [14]（只读了标题）。这是对智能体动作的人工批准，不是发布签批流程。[2][8][14]
Deploy部署◐Managed Agents run in an Anthropic-managed cloud sandbox or a self-hosted sandbox, with stateful sessions and scheduled deployments; beta, and not eligible for Zero Data Retention or HIPAA BAA coverage [1]. Usage reports expose an inference geography (global or US) and service tiers (batch, flex, priority, standard) [12]. Claude models are also served through AWS, Google Cloud and Microsoft Foundry, each setting its own retirement schedule on partner-operated platforms [11]. No serving product for a customer’s own models.Managed Agents 运行在 Anthropic 受管云沙箱或自托管沙箱中，支持有状态会话与定时部署；测试版，不符合零数据保留或 HIPAA BAA 范围 [1]。用量报表中可见推理地理区域（global 或 US）与服务层级（batch、flex、priority、standard）[12]。Claude 模型也经由 AWS、Google Cloud 与 Microsoft Foundry 提供，合作方平台自行设定退役时间表 [11]。没有面向客户自有模型的服务产品。[1][12][11]
Monitor监控◐Console session viewer for Managed Agents: transcript, tool-call counts and failures, token usage, list cost, and enforced per-session budgets [3]. Usage report API groups by workspace, API key, model, service tier and inference geography [12]. Claude Code exports OpenTelemetry metrics and events, off unless enabled, with prompt and tool content redacted by default [6]; team dashboards show usage and contribution metrics [7]. No online quality or drift monitoring found.针对 Managed Agents 的 Console 会话查看器：对话记录、工具调用次数与失败、token 用量、标价成本，以及可强制执行的单会话预算 [3]。用量报表 API 可按工作区、API 密钥、模型、服务层级与推理地理区域分组 [12]。Claude Code 导出 OpenTelemetry 指标与事件，默认关闭，开启后提示词与工具内容默认被脱敏 [6]；团队仪表板展示用量与贡献指标 [7]。没有发现线上质量或漂移监控。[3][12][6][7]
Governance治理◐Compliance API for Claude Enterprise and Console customers: Activity Feed, chats, files, projects, session transcripts (Claude Code, Cowork and others), directory and effective organisation settings, with delete operations; Compliance Access Key or Admin API key; inference hooks (beta) can deny a prompt in real time [4]. Claude Code managed settings sit at the highest precedence level [9] and admins can push telemetry settings that developer repositories cannot override [6]. No agent registry, model inventory or lineage product found, so marked partial.面向 Claude Enterprise 与 Console 客户的合规 API：活动流、对话、文件、项目、会话记录（Claude Code、Cowork 等）、目录与组织生效设置，并支持删除；使用 Compliance Access Key 或 Admin API 密钥；推理钩子（测试版）可实时拒绝某个提示词 [4]。Claude Code 托管设置处于最高优先级 [9]，管理员可下发开发者仓库无法覆盖的遥测设置 [6]。没有发现智能体注册表、模型清单或血缘产品，因此标为部分。[4][9][6]

Observations观察

Disclosure: this review was written with Claude, a model made by Anthropic. The method is the same as for the other vendors (vendor documentation only, marks only where a page supports them), but treat this page as the one most exposed to favourable reading and check it first.

披露：本次复核由 Claude 协助撰写，而 Claude 是 Anthropic 的模型。方法与其他厂商相同（只读厂商文档，仅在页面有依据时打分），但这一页最容易出现偏向性解读，建议优先复核。

Claude Managed Agents is in beta, requires the managed-agents-2026-04-01 header, and is not currently eligible for Zero Data Retention or HIPAA BAA coverage because sessions are stateful and stored server-side [1].

Claude Managed Agents 处于测试版，需要 managed-agents-2026-04-01 请求头；由于会话有状态且存储在服务端，目前不符合零数据保留或 HIPAA BAA 范围 [1]。

Code Review is not available to organisations with Zero Data Retention or the HIPAA configuration, and a review never approves or blocks a pull request [8].

启用零数据保留或 HIPAA 配置的组织不能使用 Code Review；评审从不批准或阻止拉取请求 [8]。

Claude Code telemetry is off unless enabled; prompt text, tool details and raw API bodies are redacted or not recorded by default; a repository’s own settings cannot enable telemetry, choose its destination or capture content [6].

Claude Code 遥测默认关闭；提示词文本、工具细节与原始 API 正文默认被脱敏或不记录；仓库自己的设置不能开启遥测、选择目的地或采集内容 [6]。

Method and limits: vendor documentation only, no hands-on testing. Several long pages were read only in part (permission policies, monitoring, settings, evaluations), and the documentation indices were used for page titles only; those claims are limited accordingly. 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。若干长页面只读了一部分（权限策略、监控、设置、评测），文档索引只用于核对页面标题；相关论断因此受限。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

Anthropic’s documented control surface is about what agents did and who may start them (permission policies, session transcripts, Compliance API, managed settings), not about registering and approving the agents themselves. That is a different shape from the registry-and-approval layers documented by AWS and Microsoft, and it may mean customers pair Anthropic with a hyperscaler control plane. This bears on Q2 (which systems of record own agent action logs): the Compliance API is an explicit candidate source for session transcripts.

Anthropic 文档中的控制面关注“智能体做了什么、谁能启动它们”（权限策略、会话记录、合规 API、托管设置），而不是对智能体本身做登记与审批。这与 AWS、Microsoft 文档中的“注册表加审批”层形态不同，客户可能需要把 Anthropic 与某个超大规模云的控制平面搭配使用。这关系到 Q2（谁承载智能体动作日志）：合规 API 是会话记录的一个明确候选来源。

No evaluation platform appears in the documentation index; evaluation is presented as guidance and the user’s own code. If that holds, evaluation tooling for Claude-based systems comes from partners or the open-source ecosystem (see the open-source page), consistent with evaluation tooling staying specialised (H4 on the thesis page). The evidence is an absence in an index, so it is weak.

文档索引中没有评测平台；评测被呈现为指导加用户自己的代码。如果属实，基于 Claude 的系统的评测工具要靠合作伙伴或开源生态（见开源页面），这与“评测工具保持专门化”（论点页 H4）一致。但证据只是索引中的缺席，较弱。

Strategic position战略位置

FACT. Managed Agents is beta; Claude Code, Code Review, analytics and the Compliance API are documented as available on stated plans [1][8][7][4].

FACT。Managed Agents 为测试版；Claude Code、Code Review、分析与合规 API 在文档中按所述计划提供 [1][8][7][4]。

INTERPRETATION. Strengths: a leading coding-agent product with admin controls and telemetry that repositories cannot override; a detailed Compliance API; availability on all three hyperscalers. Gaps: no registry, approval workflow or lineage; no evaluation platform found; no classical-ML products; key agent infrastructure still in beta with data-retention limits.

INTERPRETATION。优势：领先的编码智能体产品，配有管理员控制和仓库无法覆盖的遥测；详尽的合规 API；三大超大规模云均可使用。缺口：没有注册表、审批流程或血缘；没有发现评测平台；没有传统 ML 产品；关键的智能体基础设施仍为测试版且有数据保留方面的限制。

HYPOTHESIS. Anthropic will extend Managed Agents with more governance objects as it leaves beta. Evidence that would test this: Managed Agents GA notes mentioning registry, approval or evaluation features.

HYPOTHESIS。Anthropic 会在 Managed Agents 结束测试后增加更多治理对象。检验证据：Managed Agents 正式发布说明中是否出现注册表、审批或评测功能。

Open questions待解问题

Is there any first-party evaluation tooling beyond the guidance page (for example in the Console or Agent SDK)? The index showed none; check the Console and release notes.

除指导页外是否有厂商自有的评测工具（例如在 Console 或 Agent SDK 中）？索引中没有；需查 Console 与发布说明。

What do the full Managed Agents permission policies and “secure deployment” pages add (only the opening was read)?

Managed Agents 的完整权限策略与“安全部署”页面还有什么内容（只读了开头）？

How do Compliance API records of Claude Code sessions compare with OpenTelemetry events as evidence of agent actions? (Relevant to Q2 and Q6.)

合规 API 中的 Claude Code 会话记录与 OpenTelemetry 事件，哪个更适合作为智能体动作的证据？（关系到 Q2 与 Q6。）

Require and Plan are “?”. Is there any planning feature beyond model-status and migration guidance?

Require 与 Plan 仍为“?”。除模型状态与迁移指导外是否有规划功能？

References参考来源

  1. Anthropic, “Claude Managed Agents overview” (beta; header managed-agents-2026-04-01). https://platform.claude.com/docs/en/managed-agents/overview vendor documentation · S-0108
  2. Anthropic, “Permission policies,” Managed Agents (beta). Only the opening of a long page was read. https://platform.claude.com/docs/en/managed-agents/permission-policies vendor documentation (partial) · S-0109
  3. Anthropic, “Inspect sessions and track usage,” Managed Agents (beta). https://platform.claude.com/docs/en/managed-agents/session-observability vendor documentation · S-0110
  4. Anthropic, “Compliance API.” https://platform.claude.com/docs/en/manage-claude/compliance-api vendor documentation · S-0111
  5. Anthropic, Claude Code “Overview.” https://code.claude.com/docs/en/overview vendor documentation · S-0112
  6. Anthropic, Claude Code “Monitoring” (OpenTelemetry). Only the first part of a long page was read. https://code.claude.com/docs/en/monitoring-usage vendor documentation (partial) · S-0113
  7. Anthropic, “Track team usage with analytics,” Claude Code. https://code.claude.com/docs/en/analytics vendor documentation · S-0114
  8. Anthropic, “Code Review,” Claude Code. https://code.claude.com/docs/en/code-review vendor documentation · S-0115
  9. Anthropic, “Settings files and precedence,” Claude Code. Only the opening of a long page was read. https://code.claude.com/docs/en/settings vendor documentation (partial) · S-0116
  10. Anthropic, “Define success and build evaluations.” Only the first part of a long page was read. https://platform.claude.com/docs/en/test-and-evaluate/develop-tests vendor documentation (partial) · S-0117
  11. Anthropic, “Model deprecations.” https://platform.claude.com/docs/en/about-claude/model-deprecations vendor documentation · S-0118
  12. Anthropic, “Get Messages Usage Report,” API reference. https://platform.claude.com/docs/en/api/beta/organization/usage_report/retrieve_messages vendor documentation · S-0119
  13. Anthropic, Claude Developer Platform documentation index (llms.txt), used for page titles only. https://platform.claude.com/llms.txt vendor documentation index (titles only) · S-0120
  14. Anthropic, Claude Code documentation index (llms.txt), used for page titles only. https://code.claude.com/docs/llms.txt vendor documentation index (titles only) · S-0121