AWS

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

厂商档案 — AWS。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
Amazon Web Services (AWS)
Category类别
Hyperscaler: managed ML platform (Amazon SageMaker AI), model-hosting service (Amazon Bedrock) and an agent platform (Amazon Bedrock AgentCore); it hosts models from several labs (interpretation).超大规模云：托管 ML 平台（Amazon SageMaker AI）、模型托管服务（Amazon Bedrock）与智能体平台（Amazon Bedrock AgentCore）；托管多家实验室的模型（解读）。
Primary buyer主要买家
Developers and ML platform teams; AgentCore pages also address teams building an internal “agent platform” with approved tools and governed access [1]. Observation from the stated use cases.开发者与 ML 平台团队；AgentCore 页面还面向要建设内部“智能体平台”（含已批准工具与受管访问）的团队 [1]。依据页面所述用例的观察。
Relevant products相关产品
Amazon Bedrock AgentCore: Runtime, Gateway, Identity, Observability, Evaluations, Policy, Registry [1][2]; Amazon Bedrock evaluations and Guardrails [6][7]; SageMaker AI Model Registry, Model Cards, Pipelines, Model Monitor, inference endpoints [3][5][9][4][8].Amazon Bedrock AgentCore：Runtime、Gateway、Identity、Observability、Evaluations、Policy、Registry [1][2]；Amazon Bedrock 评测与 Guardrails [6][7]；SageMaker AI 的 Model Registry、Model Cards、Pipelines、Model Monitor、推理端点 [3][5][9][4][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 choice in Bedrock and JumpStart was not researched; Bedrock evaluation jobs can compare model outputs [6], which supports validation more than planning. Not marked.未研究 Bedrock 与 JumpStart 的模型选择；Bedrock 评测任务可以比较模型输出 [6]，这更偏验证而非规划。不打分。[6]
Build构建●AgentCore Harness: a managed agent loop, one API call with model, system prompt and tools, each session in an isolated microVM; works with Bedrock, OpenAI, Gemini and OpenAI-compatible models [1]. AgentCore services also work with CrewAI, LangGraph, LlamaIndex and Strands Agents [1]. Classical ML: SageMaker Pipelines (visual editor, SDK) orchestrating processing, training, fine-tuning and evaluation steps [9]. A coding-agent product was not researched; AgentCore lists “coding assistance” only as a use case [1].AgentCore Harness：受管的智能体循环，一次 API 调用指定模型、系统提示词与工具，每个会话在隔离的 microVM 中运行；可配合 Bedrock、OpenAI、Gemini 与兼容 OpenAI 的模型 [1]。AgentCore 各服务也可配合 CrewAI、LangGraph、LlamaIndex 与 Strands Agents [1]。传统 ML：SageMaker Pipelines（可视化编辑器、SDK）编排处理、训练、微调与评估步骤 [9]。未研究编码智能体产品；AgentCore 只把“编码辅助”列为用例 [1]。[1][9]
Validate验证●Bedrock evaluations: programmatic model evaluation with built-in or custom datasets, evaluation with a judge model, evaluation with human workers, and RAG evaluations of Knowledge Bases or outside RAG sources [6]. AgentCore Evaluations scores sessions, traces and spans from Strands Agents or LangGraph agents instrumented with OpenTelemetry or OpenInference, before and after deployment [1]. The SageMaker guide lists a section on evaluating, explaining and detecting bias in models (title only) [3].Bedrock 评测：用内置或自定义数据集的程序化模型评测、使用裁判模型的评测、使用人工评审的评测，以及针对 Knowledge Base 或外部 RAG 源的 RAG 评测 [6]。AgentCore Evaluations 对用 OpenTelemetry 或 OpenInference 埋点的 Strands Agents 或 LangGraph 智能体的会话、轨迹与 span 打分，覆盖上线前后 [1]。SageMaker 指南列有“评估、解释并检测模型偏差”章节（只见标题）[3]。[6][1][3]
Approve审批●SageMaker Model Registry: staging constructs for models to progress through, and management of a model’s approval status [3]; Model Cards record intended use and risk rating, with approval status updates [5]. AgentCore Registry provides “a governed workflow for publishing, reviewing, and approving resources” (agents, MCP servers, tools, skills) [1]. Marked full because both classical models and agent resources have a documented approval step; no assessment of how it is used in practice.SageMaker Model Registry：模型逐级推进的阶段结构，以及对模型审批状态的管理 [3]；Model Cards 记录预期用途与风险评级，并支持审批状态更新 [5]。AgentCore Registry 提供“发布、评审与批准资源的受管流程”（智能体、MCP 服务器、工具、技能）[1]。标为完整，是因为传统模型与智能体资源都有文档化的审批步骤；未评估其实际使用情况。[3][5][1]
Deploy部署●SageMaker inference: real-time, serverless and asynchronous endpoints; JumpStart for low-code deployment, ModelBuilder in the Python SDK, CloudFormation and CI/CD for scale [8]; Model Registry can automate deployment with CI/CD [3]. Bedrock Provisioned Throughput (hourly, 1- or 6-month commitment; required to use a customised model) [10]. AgentCore Runtime: serverless, session isolation, fast cold starts, extended runtime for asynchronous agents [1].SageMaker 推理：实时、无服务器与异步端点；JumpStart 低代码部署，Python SDK 的 ModelBuilder，规模化用 CloudFormation 与 CI/CD [8]；Model Registry 可用 CI/CD 自动部署 [3]。Bedrock Provisioned Throughput（按小时计费，可承诺 1 个月或 6 个月；使用定制模型必须购买）[10]。AgentCore Runtime：无服务器、会话隔离、快速冷启动、支持异步智能体的长时间运行 [1]。[8][3][10][1]
Monitor监控◐Agents: AgentCore Observability emits OpenTelemetry-compatible telemetry to Amazon CloudWatch (session count, latency, duration, token usage, error rates; traces and dashboards) [2]. Classical ML: Model Monitor covers data quality, model quality, bias drift and feature-attribution drift, but the page says it is no longer open to new customers and no new features are planned [4]. Marked partial because first-party model monitoring is frozen for new customers.智能体：AgentCore Observability 把兼容 OpenTelemetry 的遥测送入 Amazon CloudWatch（会话数、延迟、时长、token 用量、错误率；追踪与仪表板）[2]。传统 ML：Model Monitor 覆盖数据质量、模型质量、偏差漂移与特征归因漂移，但页面说明它不再向新客户开放，也不计划新增功能 [4]。标为部分，是因为厂商自有的模型监控对新客户已冻结。[2][4]
Governance治理●SageMaker Model Cards (intended use, risk rating, training and evaluation details, immutable versions, PDF export) integrated with Model Registry [5][3]; model lineage in Model Registry and ML Lineage Tracking in Pipelines [3][9]. Bedrock Guardrails: content, denied-topic, word and sensitive-information filters and contextual grounding checks [7]. AgentCore Identity, Policy (deterministic rules in natural language or a Cedar-compatible language, intercepting every tool call through Gateway) and Registry [1].SageMaker Model Cards（预期用途、风险评级、训练与评估详情、不可变版本、PDF 导出）与 Model Registry 集成 [5][3]；Model Registry 中的模型血缘与 Pipelines 的 ML Lineage Tracking [3][9]。Bedrock Guardrails：内容、禁用主题、词语与敏感信息过滤，以及上下文有据性检查 [7]。AgentCore Identity、Policy（用自然语言或兼容 Cedar 的语言编写的确定性规则，经 Gateway 拦截每次工具调用）与 Registry [1]。[5][3][9][7][1]

Observations观察

The SageMaker Model Monitor page states that Model Monitor is no longer open to new customers; existing customers continue as normal, AWS keeps investing in security and availability, and no new features are planned [4].

SageMaker Model Monitor 页面声明：Model Monitor 不再向新客户开放；现有客户照常使用，AWS 继续投入安全与可用性改进，但不计划新增功能 [4]。

AgentCore Registry is described as a catalogue with a governed workflow for publishing, reviewing and approving resources, and it works with resources deployed on AWS, on premises or on other clouds [1].

AgentCore Registry 被描述为带有发布、评审与批准受管流程的目录，并可用于部署在 AWS、本地或其他云上的资源 [1]。

AgentCore Evaluations currently names Strands Agents and LangGraph as the frameworks whose traces it evaluates [1].

AgentCore Evaluations 目前点名 Strands Agents 与 LangGraph 作为其评估轨迹的框架 [1]。

The overview page used here does not carry preview labels for Policy, Evaluations, Optimization or Registry; their release status was not verified (a web search surfaced 2025 preview announcements and unconfirmed GA dates). Method and limits: vendor documentation only, no hands-on testing; pages were read with the built-in browser (article text), some were cut at a character limit. Marks mean a first-party feature is documented, not that it is good or widely used.

本次所用的概览页没有给 Policy、Evaluations、Optimization、Registry 标预览；它们的发布状态没有核实（网页搜索看到 2025 年的预览公告与未经证实的 GA 日期）。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，部分在字数上限处被截断。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

AWS shows the clearest split between the two generations of tooling in the documents read: the classical ML stack has the most mature approval and documentation objects (Model Registry approval status, Model Cards with risk rating) while its monitoring is frozen for new customers, and the agent stack (AgentCore) carries the active investment in evaluation, policy and registry. This is a data point for Q1 (where classical ML and agent lifecycles diverge enough to need separate tooling) and, weakly, for H4 on the thesis page.

在已读文档中，AWS 最清楚地显示了两代工具的分野：传统 ML 栈拥有最成熟的审批与文档对象（Model Registry 审批状态、带风险评级的 Model Cards），其监控对新客户却已冻结；智能体栈（AgentCore）则承载评测、策略与注册表方面的积极投入。这是 Q1（传统 ML 与智能体的生命周期在哪里分叉到需要不同工具）的一个数据点，对论点页 H4 是弱支持。

AgentCore is built to be framework- and model-neutral (it lists OpenAI, Gemini, Claude, Llama and Mistral models and several open-source frameworks) [1], so AWS competes here on the control layer rather than on the model.

AgentCore 以框架与模型中立为设计目标（列出 OpenAI、Gemini、Claude、Llama、Mistral 模型及多种开源框架）[1]，因此 AWS 在这里竞争的是控制层，而不是模型。

Strategic position战略位置

FACT. The Bedrock evaluation, Guardrails and AgentCore pages describe first-party features at each stage marked above; Model Monitor is frozen for new customers [4].

FACT。Bedrock 评测、Guardrails 与 AgentCore 页面描述了上文各阶段的厂商自有功能；Model Monitor 对新客户已冻结 [4]。

INTERPRETATION. Strengths: documented approval and model-documentation objects for classical models; a modular, neutral agent platform with identity, policy and registry; hosting of many third-party models. Gaps: no requirements or planning stage found; monitoring for classical models is frozen; agent evaluation names only two frameworks; many names (SageMaker, SageMaker AI, SageMaker Unified Studio, Bedrock, AgentCore) make the portfolio hard to map.

INTERPRETATION。优势：传统模型有文档化的审批与模型文档对象；模块化、中立的智能体平台，带身份、策略与注册表；托管众多第三方模型。缺口：没有发现需求或规划阶段；传统模型监控已冻结；智能体评测只点名两个框架；名称繁多（SageMaker、SageMaker AI、SageMaker Unified Studio、Bedrock、AgentCore），产品组合难以梳理。

HYPOTHESIS. AWS will steer new customers of classical-ML monitoring toward partner or open-source tools while investing in agent observability. Evidence that would test this: announcements after the Model Monitor availability change, and what AWS recommends instead.

HYPOTHESIS。AWS 会引导传统 ML 监控的新客户转向合作伙伴或开源工具，同时把投入放在智能体可观测上。检验证据：Model Monitor 可用性变更之后的公告，以及 AWS 推荐的替代方案。

Open questions待解问题

What does the “Amazon SageMaker Model Monitor availability change” page say, and what does AWS recommend instead? (Only the notice on the Model Monitor page was read.)

“Amazon SageMaker Model Monitor availability change”页面写了什么，AWS 推荐什么替代？（只读了 Model Monitor 页面上的提示。）

What is the release status (preview or GA) of AgentCore Policy, Evaluations, Optimization and Registry? Check the AgentCore release notes.

AgentCore 的 Policy、Evaluations、Optimization、Registry 处于预览还是正式可用？需查 AgentCore 发布说明。

How is the AgentCore Registry approval workflow actually configured (roles, states)? Only the overview sentence was read.

AgentCore Registry 的审批流程实际如何配置（角色、状态）？只读了概览中的一句话。

Require and Plan are “?”; coding-agent products (for example Kiro) and Bedrock model-selection tooling were not researched.

Require 与 Plan 仍为“?”；编码智能体产品（如 Kiro）与 Bedrock 的模型选择工具未研究。

References参考来源

  1. AWS, “What is Amazon Bedrock AgentCore?” (Developer Guide overview). https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html vendor documentation · S-0098
  2. AWS, “Observe your agent applications on Amazon Bedrock AgentCore Observability.” https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability.html vendor documentation · S-0099
  3. AWS, “Model Registration Deployment with Model Registry,” Amazon SageMaker AI Developer Guide. https://docs.aws.amazon.com/sagemaker/latest/dg/model-registry.html vendor documentation · S-0100
  4. AWS, “Data and model quality monitoring with Amazon SageMaker Model Monitor.” https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor.html vendor documentation · S-0101
  5. AWS, “Amazon SageMaker Model Cards.” https://docs.aws.amazon.com/sagemaker/latest/dg/model-cards.html vendor documentation · S-0102
  6. AWS, “Evaluate the performance of Amazon Bedrock resources.” https://docs.aws.amazon.com/bedrock/latest/userguide/evaluation.html vendor documentation · S-0103
  7. AWS, “Detect and filter harmful content by using Amazon Bedrock Guardrails.” https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html vendor documentation · S-0104
  8. AWS, “Deploy models for inference,” Amazon SageMaker AI Developer Guide. https://docs.aws.amazon.com/sagemaker/latest/dg/deploy-model.html vendor documentation · S-0105
  9. AWS, “Pipelines,” Amazon SageMaker AI Developer Guide. https://docs.aws.amazon.com/sagemaker/latest/dg/pipelines.html vendor documentation · S-0106
  10. AWS, “Increase model invocation capacity with Provisioned Throughput in Amazon Bedrock.” https://docs.aws.amazon.com/bedrock/latest/userguide/prov-throughput.html vendor documentation · S-0107