Azure AI Foundry vs Vertex AI vs Bedrock vs Claude Platform: where should your AI agents run?
Where your agents run decides who can see your data, how they sign in and what you can audit.
Most agent pilots do not stall on the model. They stall in the security review, when someone asks where the agent runs, who can see the data and what gets logged. With no answer, the pilot waits. With the wrong answer, you re-plumb sign-in, memory and logging after the build.
This guide compares Azure AI Foundry vs Vertex AI vs Bedrock vs Claude Platform on those questions, using what each vendor's own documentation says today.
Short answer: run your agents where your data and your sign-in already live, unless a specific requirement points elsewhere: a model, a managed feature or a residency rule. The four options now are Microsoft Foundry (formerly Azure AI Foundry), Google's Gemini Enterprise Agent Platform (formerly Vertex AI), Amazon Bedrock AgentCore and Claude Platform. Below are a dated comparison table, a decision checklist and questions for each vendor.
We have no partnership with any of these vendors. We deploy where your data lives, so we have no reason to steer you.
The names changed: what you will see in the docs today
Vendor names have moved more than once recently. Older articles, SDKs and API hosts still use the old ones, so check which product a page describes.
| Name you may know | Current name | What to know |
|---|---|---|
| Azure AI Foundry | Microsoft Foundry (agent product: Foundry Agent Service) | Microsoft's current agent documentation uses the new name. [S1] |
| Vertex AI, Vertex AI Agent Engine | Gemini Enterprise Agent Platform, Agent Runtime | Google's Python SDK says the platform was formerly known as Vertex AI. [S2, S9] |
| Bedrock Agents | Bedrock Agents Classic, succeeded by Amazon Bedrock AgentCore | Closed to new customers from 30 July 2026. [S5] |
| Claude Platform | Claude Platform, also offered as Claude Platform on AWS | Anthropic operates it, with AWS sign-in and billing. It is not Bedrock. [S3, S8] |
Two cautions. First, "Bedrock Agents" is no longer an option for a new AWS account, yet many older comparisons still recommend it. If you already run it, ask AWS about the migration path.
Second, names are still moving. Re-check any product name, including those in this post, on the day you sign.
Start from where your data lives
Three questions decide most of the choice. All three concern things you already own.
Data location and processing
"Data residency" is at least three things: where data is stored, where the model runs, and which company operates the service.
The same Claude model can be run by AWS, Google, Microsoft or Anthropic, with different answers. AWS describes two paths. Amazon Bedrock keeps your data within AWS infrastructure, with regional data residency. [S4] Claude Platform on AWS is operated by Anthropic. [S8]
Microsoft's documentation for Claude in Foundry compares versions "Hosted on Azure" and "Hosted on Anthropic infrastructure". [S7] Ask which applies to the model you want.
Identity and access
Identity means how people and systems sign in, and what each may do. For agents, the question is whether each agent gets its own identity and role, or shares one key that every system trusts.
If your directory is Microsoft Entra, Foundry's documentation lists Entra identity and role-based access. [S1] AgentCore Identity integrates with providers such as Okta, Entra and Amazon Cognito. [S6] Claude Platform on AWS uses AWS IAM. [S8]
Network boundaries
Where does the agent's working state live, and can it reach your private systems? Foundry's documentation describes private networking for prompt agents, your own Azure virtual network for hosted agents, and your own storage, search and database for conversation state. [S1] For the other platforms, ask the vendor (see the list below).
The reader test: if you cannot answer these three for a platform, do not shortlist it.
If agents will call your own systems, this overlaps with LLM integration: which endpoints the agent may reach, and with whose credentials.
Azure AI Foundry vs Vertex AI vs Bedrock vs Claude Platform compared
The table shows what each vendor's documentation says. It is not a ranking. Cells we could not confirm say so.
| Dimension | Microsoft Foundry (Azure AI Foundry) | Gemini Enterprise Agent Platform (Vertex AI) | Amazon Bedrock AgentCore (Bedrock) | Claude Platform |
|---|---|---|---|---|
| Agent hosting | Foundry Agent Service: prompt agents (configuration only) and hosted agents (your code in a container). [S1] | Agent Runtime (previously Agent Engine): managed hosting and serving for agent code. [S2] | AgentCore Runtime: deploy agents built with any framework and model. [S6] | Managed Agents (beta): Anthropic runs the agent loop and a container for each session. [S8] |
| Models | Any supported model in the Foundry catalog, including Claude. [S1, S7] | Gemini plus partner and open-weight models in Model Garden, including Claude. [S2] | Any model on Runtime; Claude models through Amazon Bedrock. [S4, S6] | Claude models, also offered on Foundry, Google and AWS. [S8] |
| Identity | Microsoft Entra identity and role-based access control. [S1] | Agent Identity: credentials for each agent. [S2] | AgentCore Identity integrates with Okta, Entra and Amazon Cognito. [S6] | On AWS: SigV4 sign-in and AWS IAM access control. [S8] |
| Network and data location | Private networking for prompt agents; your own virtual network for hosted agents; your own storage for state. [S1] | Not confirmed from pages we could read. Ask which regions store Memory and Sessions. | AWS says Bedrock keeps data within AWS infrastructure, with regional data residency. [S4] | Anthropic operates it. Ask Anthropic where data is processed and stored, and about EU options. |
| Guardrails and governance | Content filters and guardrails against prompt injection. [S1] | Agent Gateway, Agent Registry and Model Armor. Ask which are generally available. [S2] | Bedrock Guardrails on the Bedrock path; AgentCore Policy uses Cedar policies. [S4, S6] | Not confirmed from pages we could read. Ask what permission and approval controls exist. |
| Logs and tracing | Tracing, metrics, evaluations and Application Insights. [S1] | Agent Observability for agent reasoning and tool calls. [S2] | AgentCore Observability, based on OpenTelemetry. [S6] | Sessions stream events. For audit export and retention, ask Anthropic. [S8] |
| Agent memory | Ask Microsoft whether built-in memory is generally available in your region. | Memory Bank for long-term memory. [S2] | AgentCore Memory: managed memory for agents. [S6] | Memory stores that sessions can attach. [S8] |
Vendor documentation checked 14 October 2026. Features marked preview or beta can change. Confirm status with the vendor before you decide.
Sources for this table
- [S1] Microsoft Learn, What is Microsoft Foundry Agent Service?
- [S2] Google, Gemini Enterprise Agent Platform samples and product overview and documentation
- [S3] AWS Machine Learning Blog, Introducing Claude Platform on AWS
- [S4] AWS What's New, Claude Opus 5.5 is now available on AWS
- [S5] AWS, Bedrock Agents Classic maintenance mode
- [S6] AWS, Amazon Bedrock AgentCore samples and developer guide
- [S7] Microsoft Learn, Claude models hosting comparison
- [S8] Anthropic, Claude API skill, Claude Platform on AWS and platform availability
- [S9] Google, python-aiplatform README
Reading the table as a buyer
- On AWS, look at AgentCore, not Bedrock Agents.
- Guardrails and logging exist on all of them, but they differ. Ask what is logged, where it goes and who can read it.
- Billing differs by path. Foundry hosted agents add container compute to per-call model and tool use. [S1] Anthropic says Claude on Foundry is billed through Microsoft Marketplace, and Claude Platform on AWS through AWS Marketplace. [S8] We list no prices. Ask for current rates.
Is Claude a fifth choice or a model on the other four?
Both. Claude is a model family, and it is also a platform.
As a model, Claude is offered on Microsoft Foundry, Google's Model Garden, Amazon Bedrock and Claude Platform on AWS, and directly from Anthropic. [S2, S4, S7, S8]
Features differ by path. Anthropic's availability table lists Managed Agents (beta) on its own API and on Claude Platform on AWS, but not on Bedrock or Google's platform; for Foundry, it marks the answer as inferred. [S8] Elsewhere you use the Claude API with tool use instead.
A decision checklist
Each "yes" points to a question to ask, not to a winner. Use it in a security or architecture review when you weigh Azure AI Foundry vs Vertex AI vs Bedrock.
| Question | If yes, ask |
|---|---|
| Is your identity provider Microsoft Entra, and are most users in Microsoft 365? | What does Foundry's identity and role-based access cover for each agent? |
| Is your main data store on Google Cloud? | Which Agent Platform features are generally available in my region? |
| Do your workloads and logging already run on AWS? | For AgentCore or Claude Platform on AWS: who processes the data? |
| Must processing stay in the EU? | Where is the model hosted, where is state stored, and who operates it? |
| Do you need one specific model? | Is that model on the platform, with which features? |
| Must agents reach tools in a private network? | Does the runtime support private networking in my setup? |
| Will auditors ask what the agent did? | What is logged, where, and can we export it? |
Questions to ask each vendor before you commit
- Where is each Claude model in your catalog hosted, and who operates it?
- Which agent features are generally available in my region today, and which are preview?
- Can each agent get its own identity and role?
- What exactly is logged, where, and for how long?
- How is agent memory stored, and can we delete it?
- What changes if a product is renamed or retired, and how are we told?
Not sure which answers apply to your company? Book an AI scoping call. Bring your cloud setup and the workflow you want an agent to run. You leave with the platform questions that matter for you, and a view on whether a scoping sprint is worth doing.
How we would approach this for your company
We start with the workflow, not the platform.
- AI Scoping Sprint. We map the systems and data the agent touches, check where that data may be processed, and work through the checklist above against current documentation. You get a written recommendation and build plan, with a cost model for building and running it. Your team can act on it with or without us.
- Pilot build. One workflow on the platform the sprint pointed to, connected to real systems in staging or limited production, with success criteria and an evaluation set agreed up front.
- Production build. Architecture, integrations, evaluation, security review steps, monitoring and handover. This is where your agent systems go live in your own cloud account.
- Run and improve (optional). Monitoring, evaluation runs, model and provider updates, and cost control.
The sprint ends in a go or no-go and a platform decision you can defend to security and legal. For a rough running-cost view first, use the AI agent cost calculator. The stages are in how we work.
What stays the same on any platform
The design around the model does not change. We call it the AI agent harness: tools and permissions, memory, logging, quality gates, human approval and cost caps.
Only the hosting, identity and logging plumbing changes. Keep prompts, tool definitions and evaluation sets in your own repository, and a later switch is mostly re-plumbing, not a rewrite. That is our engineering judgment, not a vendor statement.
We design, build and run agent systems, and we deploy where your data lives: Microsoft Foundry, Google's agent platform, AWS or Claude Platform. That describes what we do, not a partnership or certification. We run our own team of agents on Claude Managed Agents, with human oversight; see how we run ByteandRise. We work under NDA.
Settle where your agents run before you build them. In an AI scoping call, we review your data, your sign-in setup and the workflow, and name the decisions a sprint should close. Book an AI scoping call.
FAQ
Is Azure AI Foundry the same as Microsoft Foundry?
Yes. Microsoft renamed Azure AI Foundry to Microsoft Foundry, and its agent documentation now uses the new name. Older articles, SDKs and job posts still say Azure AI Foundry, so search for both names when you research or review a vendor's quote.
Is Vertex AI still a product?
Google now calls it Gemini Enterprise Agent Platform (formerly Vertex AI), and Vertex AI Agent Engine is now Agent Runtime. Google's Python SDK still carries the old package name. Names are still moving, so check the release notes before you put one in a contract.
Should we build on Bedrock Agents or AgentCore?
For a new AWS account, AgentCore. AWS renamed Bedrock Agents to Bedrock Agents Classic and closed it to new customers on 30 July 2026. If you already use it, ask AWS about the migration path and what support continues, and plan that move before you add new agents to it.
Should we use Claude through our cloud or on Claude Platform directly?
It depends on data location and features. Claude models are offered on Microsoft Foundry, Google's platform and AWS, and Claude Platform on AWS is a separate path from Bedrock. Anthropic's availability table lists Managed Agents (beta) only on its own API and Claude Platform on AWS. Ask each vendor who hosts the model and which features you get.
How hard is it to switch platforms later?
Easier if you keep prompts, tool definitions, evaluation sets and your AI agent harness design in your own repository. Memory, identity and logging are the parts you would re-plumb. This is our engineering judgment, not a vendor statement, so test it in a pilot before you rely on it.