- Home
- AI agents
- Platforms and frameworks
- Microsoft vs AWS vs Google
Build · Platforms and frameworks
Microsoft vs AWS vs Google for AI agents: capability by capability
Microsoft, AWS and Google each sell an agent platform with the same nine parts: build tools, runtime, memory, identity, registry, gateway, guardrails, observability and evaluation. They differ in who they are built for, where agents are published and which parts were still preview on 2 Nov 2026. This page lists each part with its status and date, without ranking.
In this guide: 83 pages
Key takeaways
- Microsoft spans low-code and pro-code in one tenant and publishes agents into Microsoft 365; AWS sells modular services that work with any framework and model; Google pairs an open-source development kit with a managed runtime and an employee-facing app. Those are design choices, and none of them is a ranking.
- By 2 Nov 2026 the governance layer was generally available on all three: Agent 365 and Entra Agent ID (1 May 2026), AgentCore Policy (3 Mar 2026) and Evaluations (31 Mar 2026), Google Agent Gateway, Registry and Observability (18 Jun 2026). Google Agent Identity was still a preview API.
- The three share MCP for tools, A2A for agent-to-agent calls and OpenTelemetry for traces, and all three host the open frameworks. The portable parts of an agent are the ones you keep in your own code.
- The fit is decided by data gravity and the identity provider: Entra, AWS IAM or Google IAM. A company with data in two clouds runs two, with one registry of record per cloud.
How is each cloud’s agent stack organised?
Microsoft organises around the tenant, AWS around modular services and Google around four pillars, and the organisation tells you who each platform expects to build. The differences below are about approach, with the statuses as the vendor pages showed them on 2 Nov 2026.
Microsoft puts two build tools under one identity system. Copilot Studio serves makers with three harnesses and publishes into Teams and Microsoft 365 Copilot; Microsoft Foundry serves engineers with prompt agents, voice-based prompt agents and hosted agents that accept Agent Framework, LangGraph, the OpenAI Agents SDK, the Anthropic Agent SDK or custom code as a container. Microsoft Agent 365, generally available since 1 May 2026, is the control plane across both, with the registry, Entra access controls, Purview data security and Defender threat protection. The agent service was renamed from Azure AI Foundry at Ignite on 18 Nov 2025.
AWS sells Amazon Bedrock AgentCore as a set of services you use together or alone: Harness, Runtime, Memory, Gateway, Identity, Code Interpreter, Browser, Observability, Evaluations, Optimization, Policy, Payments and AWS Agent Registry. Runtime works with CrewAI, LangGraph, LlamaIndex, Google ADK, the OpenAI Agents SDK and Strands, with any model in or outside Bedrock. The earlier product, Amazon Bedrock Agents, became "Classic" and closed to new customers on 30 Jul 2026. Amazon Quick is the business-user workspace at the other end of the range.
Google renamed Vertex AI to Gemini Enterprise Agent Platform on 22 Apr 2026 and organises it as Build (Agent Development Kit, Agent Studio, Agent Garden, Managed Agents API), Scale (Agent Runtime, Sessions, Memory Bank, Code Execution), Govern (Agent Registry, Agent Identity, Agent Gateway, Governance Policies, Model Armor) and Optimize (evaluation, simulation, observability). Gemini Enterprise, launched 9 Oct 2025 succeeding Agentspace, is the employee front door with a no-code Agent Designer.
Design choices on each cloud, with no ranking implied.
| Dimension | Microsoft | AWS | |
|---|---|---|---|
| Primary builder | Makers in Copilot Studio and engineers in Foundry, one tenant | Engineering teams; business users in Amazon Quick | Engineers with ADK; business users in Gemini Enterprise |
| Frameworks accepted | Agent Framework, LangGraph, OpenAI Agents SDK, Anthropic Agent SDK, Copilot SDK, custom code | CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI Agents SDK, Strands | ADK in Python (2.0 generally available 19 May 2026), TypeScript, Go, Java |
| Where agents are published | Teams, Microsoft 365 Copilot, websites, phone numbers, custom apps | Your own channels through Runtime endpoints; Gateway and the registry MCP endpoint for other agents | Gemini Enterprise app, Workspace, custom apps |
| Identity model | Entra Agent ID with owners and sponsors; acts on behalf of a user or as itself | AgentCore Identity workload identities; inbound JWT, outbound OAuth and API keys; any IdP | Agent Identity with a SPIFFE ID per agent; API in preview |
| Policy point for tool calls | Entra Conditional Access for agents; API Management AI gateway; Foundry Toolboxes (preview) | AgentCore Policy on AgentCore Gateway, every call checked | Agent Gateway with IAM Unified Access Policies and Model Armor |
| Other clouds | Agent 365 registry sync with AWS Bedrock and Google (preview); Defender discovers agents on other platforms | Registry indexes agents built elsewhere by hand; auto-detection is AgentCore only | Manual registration of external A2A agents and MCP servers |
What is the same on all three?
All three run the same protocols at the edges and host the same open frameworks, so the portable part of an agent is larger than the marketing suggests. Four things carry over between clouds.
- MCP for tools. The Model Context Protocol (current specification 2026-07-28) is how agents call tools on every platform: AgentCore Gateway, Foundry Toolboxes and remote MCP servers, Google Agent Gateway. An MCP server written once attaches to all three.
- A2A for agent-to-agent calls. Version 1.0.0 (12 Mar 2026) with signed agent cards is generally available in Foundry; Copilot Studio A2A has been generally available since April 2026, though its documentation still shows 0.3-style payloads; Google Agent Registry stores v1 agent cards; the documented AgentCore path still pins the 0.3 SDK.
- OpenTelemetry for traces. AgentCore Observability emits OpenTelemetry-compatible telemetry, Foundry tracing and Google Agent Observability consume it, and the generative AI semantic conventions are still in development, so expect attribute names to change.
- Open frameworks and consumption billing. LangGraph, the OpenAI Agents SDK and Google ADK run on more than one platform, and every platform bills by use: credits, per-second compute, per invocation. The cost shape is the same even where the meters differ.
The parts that do not carry over are the managed ones: memory stores, registry records, policy rules and identity bindings. That is the lock-in to plan for, and the sub-page on build vs buy covers the exit.
When does each one fit?
Each fits the workload whose data, identities and users already live there; the table says which situation points where, and a company with data in two clouds should expect two answers.
Situations and the platform each one usually points to. None of these is a ranking.
| Situation | Usually points to | Why |
|---|---|---|
| Agents that serve staff inside Microsoft 365, built by business teams | Copilot Studio with Agent 365 | Connectors, Teams and Microsoft 365 Copilot publishing, Entra Agent ID per agent since July 2026, Purview audit |
| Custom code on Azure with a chosen model and a virtual network | Microsoft Foundry hosted agents | Any framework in a container, dedicated Entra identity, bring-your-own VNet, tracing generally available |
| Data and APIs already in AWS accounts, any framework | Amazon Bedrock AgentCore | Runtime isolation per session, Gateway and Policy on every tool call, AgentCore Identity with your existing IdP, registry across AWS Organizations |
| Teams standardised on Google Cloud and Workspace | Gemini Enterprise Agent Platform | ADK on Agent Runtime, Agent Gateway with Model Armor, Agent Registry with A2A v1, publishing into Gemini Enterprise |
| Agents that act only inside one SaaS system of record | That system’s own agents | Agentforce and ServiceNow AI Agents run within the permissions and workflows already there; cross-system work still needs a platform |
| Data in two clouds | Two platforms, one registry of record per cloud | Cross-cloud registry sync was preview on 2 Nov 2026, so plan a manual reconciliation |
Preview status matters for every row. Google Agent Identity, the Managed Agents API, Foundry Toolboxes and memory, Copilot Studio multistage approvals and the Agent 365 cross-cloud sync all carried preview labels on 2 Nov 2026 and run without production service-level agreements.
How do Microsoft, AWS and Google implement it?
The nine parts, with the product, status and date each cloud showed on 2 Nov 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Build | Copilot Studio; Foundry prompt agents and hosted agents; Microsoft Agent Framework 1.0 (3 Apr 2026) | AgentCore harness (17 Jun 2026) or any framework; Strands Agents 1.0 (15 Jul 2025) | Agent Development Kit; Agent Studio; Managed Agents API (preview, 19 May 2026) |
| Runtime | Foundry hosted agents (generally available per the 2 Jun 2026 Azure update); Copilot Studio harnesses | AgentCore Runtime (13 Oct 2025); microVM per session | Agent Runtime, formerly Vertex AI Agent Engine (renamed 22 Apr 2026) |
| Memory | Foundry memory (public preview) | AgentCore Memory (generally available; episodic strategy since Dec 2025) | Agent Platform Memory Bank; memory profiles generally available 15 Jul 2026 |
| Identity | Entra Agent ID (generally available, Learn page dated 1 May 2026) | AgentCore Identity (13 Oct 2025) | Agent Identity, SPIFFE-based (API in preview since 18 Jun 2026) |
| Registry | Agent 365 registry (1 May 2026); cross-cloud sync in preview | AWS Agent Registry (31 Aug 2026) | Agent Registry (18 Jun 2026) |
| Gateway and policy | Foundry Toolboxes (preview); API Management AI gateway; Entra Conditional Access for agents | AgentCore Gateway (13 Oct 2025); AgentCore Policy (3 Mar 2026) | Agent Gateway (18 Jun 2026); IAM Unified Access Policies (31 Aug 2026) |
| Guardrails | Prompt Shields (generally available since Aug 2024); Foundry tool-call guardrails (preview) | Bedrock Guardrails; Automated Reasoning checks (6 Aug 2025) | Model Armor; Agent Gateway integration (24 Jun 2026) |
| Observability | Foundry tracing (generally available for prompt and hosted agents); Application Insights Agents view (preview) | AgentCore Observability on CloudWatch (13 Oct 2025) | Agent Observability (18 Jun 2026) |
| Evaluation | Foundry evaluators (several preview); Copilot Studio evaluations (Mar 2026) | AgentCore Evaluations (31 Mar 2026), 13 built-in evaluators | Agent evaluation and simulation |
Dates without a label are general availability dates. Google’s release notes carry conflicting labels for Agent Identity and Microsoft’s pages disagreed on the hosted-agents date, so both are shown at the more cautious reading.
Frequently asked questions
Which cloud is best for AI agents?
None of the three is the right answer in general, and this page does not rank them. Each fits workloads whose data, identities and users already sit on it. A company with data in Azure and AWS needs two platforms and one registry of record per cloud.
Can an agent built on one cloud call tools on another?
Yes, through MCP. A tool exposed as an MCP server behind AgentCore Gateway, a Foundry Toolbox or Google Agent Gateway can be called by an agent on another platform, subject to the identity and policy each gateway enforces. The registry records and policy rules stay with the cloud that hosts the tool.
Is Amazon Bedrock AgentCore the same as Amazon Bedrock Agents Classic?
No. Amazon Bedrock Agents became "Classic" and closed to new customers on 30 Jul 2026, with its model catalog frozen. AgentCore is the modular platform AWS recommends instead, generally available since 13 Oct 2025, and AWS says some Classic features such as routing-mode multi-agent need custom framework code on the new harness.
Is Gemini Enterprise Agent Platform the renamed Vertex AI?
It is the renamed product. Google renamed Vertex AI to Gemini Enterprise Agent Platform on 22 Apr 2026, with Vertex AI Agent Engine becoming Agent Runtime, and describes the platform as an evolution of Vertex AI. The release notes and old URLs redirect to the new names.
More in Platforms and frameworks
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- What is Microsoft Foundry Agent Service?
- Microsoft Agent 365 overview
- Microsoft Agent 365 now generally available
- What’s new in Microsoft Entra Agent ID
- Harnesses in Copilot Studio
- Azure AI Services is now Foundry Tools (Azure update AZ526132)
- Guardrails and controls in Microsoft Foundry
- Set up tracing for agents in Microsoft Foundry
- What is Amazon Bedrock AgentCore?
- AgentCore Identity
- Amazon Bedrock AgentCore is now generally available
- Policy for Amazon Bedrock AgentCore generally available
- AgentCore Evaluations generally available
- AWS Agent Registry now generally available
- Amazon Bedrock Agents Classic maintenance mode
- Automated Reasoning checks in Amazon Bedrock Guardrails generally available
- Top announcements of the AWS Summit in New York 2026
- Gemini Enterprise Agent Platform overview
- Gemini Enterprise Agent Platform release notes
- Welcome to Google Cloud Next ’26
- Introducing Gemini Enterprise
- Model Context Protocol specification
- A2A protocol specification, version 1.0.0
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner