Part of
Microsoft Foundry
Guides

Amazon Bedrock vs Microsoft Foundry: Choosing Your Enterprise AI Agent Platform

Dark card: Bedrock vs Foundry, choose where your data lives; three rows, Azure SQL: Microsoft Foundry, Amazon S3: Amazon Bedrock, and Both clouds: Platform-neutral layer, highlighted in orange.

Amazon Bedrock and Microsoft Foundry both run enterprise agents. Microsoft Learn lists Azure AI Studio and Azure AI Foundry as Microsoft Foundry's former names. The platform you pick decides, for years, how your agents connect to other systems, how they are secured and how fast your team builds. The decision rarely hinges on which platform hosts the single best model. It turns on how each platform tracks an agent's work through long tasks and how it coordinates several agents in one workflow. It also turns on how tightly it integrates with your existing data estate.

Choose the platform where your data already lives. If your organization runs on SharePoint and Azure SQL, Microsoft Foundry sits close to that data. That cuts the cost of moving data out of its cloud and makes permissions simpler to set. If your data sits in Amazon S3 or Aurora, Amazon Bedrock avoids the delay and cost of fetching data from another cloud. If your workloads span both clouds, a platform-neutral orchestration layer can sit above them. That layer is software that runs your workflows across both clouds and sends each task to the model or service that fits it best. No single vendor's tools then own your architecture.

How the two platforms differ: Amazon Bedrock vs Microsoft Foundry

Core philosophy: modular APIs vs a unified workspace

Amazon Bedrock works as a modular parts catalog. AWS describes it as a fully managed service that provides secure, enterprise-grade access to high-performing foundation models from leading AI companies (Amazon Bedrock overview). You pick the model that fits the job and call it through the Bedrock APIs, with no infrastructure to run. This suits teams that want to connect pieces in their own code without committing to one vendor's set of tools. Microsoft Foundry takes the opposite approach. It groups agents, models and tools in one place, with one set of role-based access control, networking and policies (What is Microsoft Foundry?). Your models, data connections and deployments all sit under one Azure roof. Neither approach is better in itself. The right fit depends on whether your team builds better by picking parts from a catalog or by working in one unified workspace.

Which models each platform offers for agents

On either platform, an agent works best when each step uses a model suited to it. A careful reasoning step differs in speed and cost from a fast call that pulls out a field or sorts a document. The Amazon Bedrock overview, read on 11 November 2026, lists 100+ foundation models from Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI and xAI. Its Converse API takes the model as one setting in the call. Swapping a model therefore does not mean rewriting the code that runs the agent's steps. Microsoft Learn, read the same day, puts the Foundry catalog at more than 10,000 models from Microsoft, OpenAI, Anthropic, Meta and others (What is Microsoft Foundry?). Availability differs by region on both. Confirm the models you need in the regions you deploy to before you commit to an architecture.

Orchestration: how each platform runs agents

An agent needs orchestration, the code that runs its steps in order. That code has to track a long task, keep memory and call tools, and keep working when agents call each other over and over. On AWS, new agents are built on Amazon Bedrock AgentCore. AWS has closed sign-ups for the earlier Amazon Bedrock Agents, now called Agents Classic, and existing users can keep running it (Amazon Bedrock agents documentation). AgentCore runs agents serverless, so you manage no servers. It provides services for runtime, memory, identity and observability, plus a gateway that turns your APIs into MCP tools that agents can call. It works with any framework and any foundation model, including open-source frameworks such as CrewAI, LangGraph, LlamaIndex and Strands Agents (What is Amazon Bedrock AgentCore?). On Azure, Foundry Agent Service offers two kinds of agent. You define prompt agents with instructions, a model and tools, and Foundry hosts and runs them. Hosted agents run your own code, built with Microsoft Agent Framework, LangGraph or Semantic Kernel (Microsoft Learn).

Enterprise deployment: security, compliance and cost

Security boundaries and data isolation

Enterprise agents read your company's private data, so the real risk sits in the security around them. For Bedrock, AWS recommends IAM or IAM Identity Center to control access and documents Amazon VPC and AWS PrivateLink for protecting your data. It also states that model providers have no access to Bedrock logs or to customer prompts and completions (AWS, Data protection). For Foundry, Microsoft Entra identity, role-based access control, content filters, network isolation and Azure Policy apply across every project (Microsoft Learn). Compliance certifications and data isolation terms for agent workloads vary by region, service tier and configuration. Check the standards you need (SOC, ISO, HIPAA and regional equivalents) with each provider before you finalize the design.

Key evaluation criteria for a cloud AI platform comparison

When you score Bedrock and Foundry against your enterprise requirements, four criteria drive the most important trade-offs:

  • Ecosystem gravity: Where do most of your data systems live? The platform next to them avoids fetching data from another cloud. That saves the delay, the egress cost (the fee for moving data out of a cloud) and a second set of permissions to manage.
  • Orchestration complexity: Decide whether your team needs low-code tools to build agent workflows or prefers to build custom agent loops with developer SDKs. The answer tells you which platform's built-in features save time and which only add layers to work around.
  • Vendor lock-in risk: Check how easily your agent workflows could move to another platform. Proprietary action groups, native memory stores and platform-specific tool registries make that move harder. A middleware layer between your workflows and the platform can reduce the risk.
  • Cost predictability: Model use is billed by the token, a small chunk of text. An agent that runs on its own uses more tokens the more steps it takes and the more documents it pulls in. Measure a baseline cost in a controlled pilot before you roll out to production.

Cost predictability in autonomous loops

Agent costs run away when agents call other agents over and over. They also run away in retrieval-augmented generation (RAG) loops, where an agent pulls many documents into each answer. Each multiplies the tokens one request uses. Both platforms bill by consumption. AWS gives AgentCore consumption-based pricing with no upfront commitments or minimum fees (What is Amazon Bedrock AgentCore?). On Foundry, you pay as you go on your Azure bill for model tokens, hosted-agent compute and each tool you use. Microsoft charges no platform fee, as our Microsoft Foundry page sets out from Microsoft's pricing pages. The two are hard to compare without running a representative workload, and both vendors change their pricing. Before any production deployment, set hard limits: a maximum number of loops, a token budget and a circuit breaker that stops runaway repeated calls. Use the pilot to set your baseline cost.

Amazon BedrockMicrosoft Foundry
ApproachA modular parts catalogA unified workspace
ModelsFoundation models from Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI and xAIModels from Microsoft, OpenAI, Anthropic, Meta and others
AgentsAmazon Bedrock AgentCoreFoundry Agent Service
FrameworksAny framework and any foundation model, open-source frameworks such as CrewAI, LangGraph, LlamaIndex and Strands Agents among themHosted agents that run your own code, built with Microsoft Agent Framework, LangGraph or Semantic Kernel
AccessIAM or IAM Identity CenterMicrosoft Entra identity, role-based access control
NetworkAmazon VPC and AWS PrivateLinkNetwork isolation and Azure Policy
PricingAWS gives AgentCore consumption-based pricing with no upfront commitments or minimum feesPay as you go on your Azure bill for model tokens, hosted-agent compute and each tool you use
Data estateIf your data sits in Amazon S3 or AuroraIf your organization runs on SharePoint and Azure SQL

The multi-cloud alternative: one orchestration layer above both clouds

Choosing between AWS and Azure for AI agents does not have to tie your architecture to one vendor's tools. A platform-neutral orchestration layer sits above the cloud infrastructure. It sends each task to the model or service that fits it best, whichever cloud hosts it. You can use Bedrock's flexible APIs for some workloads and Foundry's enterprise data integrations for others. You do not rebuild your workflow logic every time a vendor changes its own tools. Ovidius builds this layer on n8n, self-hosted on your AWS, Azure or GCP account. On a Foundry build, n8n is the orchestration layer that moves Foundry outputs into your operational systems.

Amazon Bedrock or Microsoft Foundry
Which platform fits
Where your data lives
Workflows must stay portable
The post points to
Microsoft Foundry
Why
Built on
What each choice gives
  • SharePoint and Azure SQL: Microsoft Foundry, on Foundry Agent Service, with prompt agents or hosted agents that run your own code. Microsoft Foundry offers native data proximity that reduces egress friction and simplifies permissioning.
  • Amazon S3 or Aurora: Amazon Bedrock, on Amazon Bedrock AgentCore, which works with any framework and any foundation model. Amazon Bedrock minimizes the latency and cost penalties of cross-cloud retrieval.
  • Both clouds: A platform-neutral orchestration layer, on n8n, self-hosted on your AWS, Azure or GCP account. It routes each task to the best-fit model or service, so you can run some workloads on Bedrock and others on Foundry without rebuilding your orchestration logic when a vendor changes its tooling.
  • Workflows must stay portable: Proprietary action groups, native memory stores, and platform-specific tool registries create migration friction; middleware layers can reduce that exposure.
Ecosystem gravity and vendor lock-in risk, from the four criteria in this post.

How to decide between Bedrock and Foundry

Start with where your data lives. Score both platforms on the four criteria above. Then run a pilot on one real workflow that sets its baseline cost before you commit to production. Whichever you pick, put the limits and the cost baseline in place before the first agent reaches live records.

Frequently asked questions

Should a company on SharePoint and Azure SQL build its agents on Microsoft Foundry?

Yes, on data location alone. With data in SharePoint and Azure SQL, Foundry sits close to it. That cuts the cost of moving data out of its cloud and makes permissions simpler. Data in Amazon S3 or Aurora points to Bedrock, which avoids the delay and cost of fetching data from another cloud. Before you commit, also score orchestration complexity, vendor lock-in risk and cost predictability. Then run a pilot that measures the cost of one real workflow.

Is Azure AI Foundry the same platform as Microsoft Foundry?

Yes. Microsoft Learn lists Azure AI Studio and Azure AI Foundry as the previous brand names of the platform now called Microsoft Foundry. Guides and search results that still say Azure AI Foundry describe the same platform this post compares with Amazon Bedrock.

How do you build an agent on Amazon Bedrock today?

You build it on Amazon Bedrock AgentCore. AWS has closed sign-ups for the earlier Amazon Bedrock Agents, now called Agents Classic, though existing users can keep running it. AgentCore runs agents serverless, so there are no servers to manage. It provides runtime, memory, identity and observability, plus a gateway that turns your APIs into MCP tools. It works with any foundation model and with frameworks such as LangGraph, CrewAI, LlamaIndex and Strands Agents.

Which platform fits if our data sits on both AWS and Azure?

A platform-neutral orchestration layer above both clouds fits that case. It sends each task to the model or service that fits it best. You can run some workloads on Bedrock and others on Foundry, and you do not rebuild your workflow logic when a vendor changes its tools. Ovidius builds that layer on n8n, self-hosted on your AWS, Azure or GCP account.

How do Bedrock and Foundry charge for running agents?

Both bill by consumption. AWS gives AgentCore consumption-based pricing with no upfront commitments or minimum fees. On Foundry, you pay as you go on your Azure bill for model tokens, hosted-agent compute and each tool the agent uses. Microsoft charges no platform fee. On both, token use multiplies when agents call each other over and over or pull in many documents. Measure the cost in a pilot and set loop limits and token budgets before production.

Can Bedrock and Foundry run agents inside a private network?

Both document network controls. AWS documents Amazon VPC and AWS PrivateLink for protecting Bedrock data and recommends IAM or IAM Identity Center for access. It also keeps model providers away from customer prompts and completions. Foundry applies Microsoft Entra identity, role-based access control, network isolation and Azure Policy across every project. Certifications and isolation terms still vary by region and service tier. Confirm the standards you need with each provider.

Choosing a platform for your first enterprise agent?

To review your current cloud setup and design an agent architecture that is not locked to one vendor, schedule a consultation. A senior partner runs the call, and we pick the platform after we understand the workflow.

More on this topic

Put your first workflow on the board

Answer a few questions about your team and what you want automated, then a senior partner scopes it with you on the call.