Microsoft Foundry vs AWS AgentCore vs Claude Managed Agents
Microsoft Foundry vs AWS AgentCore vs Claude Managed Agents (May 2026)
The 2026 enterprise agent platform race has settled into three serious contenders, each anchored to a single cloud or model vendor. Microsoft Foundry (Azure-native, multi-model), Amazon Bedrock AgentCore (AWS-native, framework-agnostic), and Claude Managed Agents (Anthropic-native, opinionated). Here’s how to choose by stack, by model preference, and by ops burden.
Last verified: May 25, 2026.
TL;DR table
| Microsoft Foundry | AWS AgentCore | Claude Managed Agents | |
|---|---|---|---|
| Cloud | Azure | AWS | Anthropic (cloud-agnostic but Anthropic-only) |
| Model selection | Multi-model (OpenAI, Anthropic, Mistral, Microsoft) | Multi-model (Bedrock catalog) | Claude only |
| Framework | Microsoft Agent Framework, Copilot Studio, Foundry Agents | Any (LangGraph, CrewAI, Claude SDK, custom) | Anthropic-native |
| Identity | Entra agent identity | AgentCore Identity (IAM-native) | Anthropic OAuth + tools |
| Memory | Foundry memory primitives | AgentCore Memory (episodic) | Built-in + Dreaming research preview |
| Multi-agent | Drag-and-drop orchestration + Copilot Studio | Framework-determined | Public beta (May 6, 2026) with shared FS |
| Eval / safety | M365 Copilot Evaluations + RAMPART/Clarity | AgentCore Evaluations + Policy | Outcomes (rubric-driven, public beta) |
| Pricing model | Per-token + Foundry Hosted Agents (scale-to-zero) | Per-token + per-component runtime | API token + managed-runtime markup |
| M365 integration | Native | Via connectors | Via connectors |
| Best for | Azure + M365 shops | AWS-native, framework freedom | Anthropic-only, lowest ops burden |
The three vendor strategies
Microsoft Foundry — “Azure unified everything”
Microsoft’s pitch: one portal, one runtime, one identity model, one observability story across every model — OpenAI, Anthropic, Mistral, Microsoft. The April 22, 2026 Foundry Hosted Agents refresh added hypervisor isolation, Entra agent identity, and scale-to-zero pricing.
Strongest where:
- Azure is the primary cloud
- M365 Copilot is already deployed
- Compliance + governance is a board-level priority
- Multi-model flexibility matters
Weakest where:
- Workload is AWS-native (the integration is doable but you give up the native ops benefits)
- Teams want framework freedom outside Microsoft Agent Framework / Copilot Studio
AWS Bedrock AgentCore — “Open runtime, bring your model and framework”
AWS’s pitch: don’t pick the framework, don’t pick the model, just use AgentCore for the parts everyone has to build — runtime, memory, identity, gateway, observability. You bring LangGraph or CrewAI or AutoGen or Claude Agent SDK or your custom harness, and AgentCore handles the production primitives.
Strongest where:
- AWS is the primary cloud
- You want full framework choice (open-source frameworks especially)
- Multi-model is required (Bedrock catalog gives Anthropic, Cohere, AI21, Meta, plus Amazon Nova)
- You need fine-grained IAM-native authorization
Weakest where:
- You want a single end-to-end opinionated platform (Foundry or Claude Managed Agents are more opinionated)
- M365 integration is a top priority
Claude Managed Agents — “Anthropic-only, hosted, opinionated”
Anthropic’s pitch: if your model choice is Claude, why operate the runtime yourself? Managed Agents handles secure sandboxing, long-running sessions, scoped permissions, tool execution, and tracing. Outcomes (rubric-driven quality gates) and multi-agent orchestration are public beta. Dreaming (memory consolidation) is research preview.
Strongest where:
- You’re committed to Claude
- You want the lowest ops burden of the three
- Outcomes-style evaluation-driven development matches your workflow
- Multi-agent orchestration with shared filesystem semantics is appealing
Weakest where:
- You want multi-model flexibility
- You’re tied to AWS or Azure with strict data-residency-in-our-cloud requirements
- You want bring-your-own-framework freedom
Real-world decision framework
Pick Microsoft Foundry when:
- Azure is your primary cloud, or M365 Copilot is core to your productivity strategy.
- You want a unified portal experience across multiple models.
- Governance, Entra-based agent identity, and enterprise compliance are top priorities.
- You’re okay being inside Microsoft Agent Framework / Copilot Studio / Foundry Agents (or willing to compose them).
Pick AWS AgentCore when:
- AWS is your primary cloud.
- You want framework freedom — LangGraph, CrewAI, AutoGen, Claude Agent SDK, or your own.
- You need fine-grained IAM-native authorization for tools and data.
- You want to mix models from the Bedrock catalog.
Pick Claude Managed Agents when:
- You’re a Claude-first shop.
- You want the lowest possible ops burden.
- Outcomes (rubric-driven iteration) matches your team’s preferred dev pattern.
- You’re early enough in the journey that you can lean into Anthropic’s opinionated patterns rather than retrofitting your existing framework.
Pick none of these (run your own) when:
- True multi-cloud portability is a hard requirement.
- You need on-prem or air-gapped deployment.
- Your security model requires inspecting and controlling every runtime primitive.
- You have the team capacity to run a Kubernetes-based agent platform yourself.
What about Camunda ProcessOS, Salesforce Agentforce, and others?
These are at different layers. Foundry / AgentCore / Claude Managed Agents are infrastructure-and-runtime. ProcessOS is the process-intelligence + orchestration layer on top (and runs on AWS AgentCore). Salesforce Agentforce is the CRM-bound vertical agent layer. ServiceNow AI Agents is the ITSM-bound vertical agent layer. Most enterprises end up running:
- One infrastructure runtime (Foundry / AgentCore / Claude Managed)
- One or more vertical agent layers (Agentforce, ServiceNow, Workday Illuminate)
- Sometimes a process orchestration layer (ProcessOS, Workato, Power Automate)
- And end-user agents (M365 Copilot, Google Workspace Studio, ChatGPT Workspace)
The infrastructure decision drives the cloud bill and the ops burden; the vertical layers drive the user-facing experience.
Pricing posture (rough order)
- Cheapest at scale: AgentCore — pay per component, no minimums.
- Most predictable: Microsoft Foundry — bundled into M365 Copilot licensing for in-org workflows; pay-go for net-new.
- Lowest ops cost: Claude Managed Agents — managed runtime markup is the price for Anthropic operating it for you.
Real cost differences are dominated by token spend on the underlying model. Runtime markups are typically 10-25% of the model bill. Choose by ops fit and stack alignment, not by line-item runtime pricing.
What’s missing from each platform in May 2026
- Microsoft Foundry: Cross-cloud parity story is thin. Strong inside Azure, awkward outside.
- AWS AgentCore: Opinionated end-to-end developer experience is weaker than Foundry or Claude Managed — you assemble pieces.
- Claude Managed Agents: Anthropic-only. No path to mix in OpenAI or Gemini models inside the managed runtime.
All three are racing to add what’s missing. Expect more multi-cloud and multi-model bridges by Q4 2026.
Verdict
- Best for Azure + M365 shops: Microsoft Foundry.
- Best for AWS-native, framework freedom: AWS Bedrock AgentCore.
- Best for Claude-first, lowest ops burden: Claude Managed Agents.
- Best for true cloud portability: None of the three — run your own.
- Most common 2026 outcome: Dual stack (Foundry + AgentCore) with consistent org-level governance.
The May 2026 enterprise agent platform decision is fundamentally a cloud and model decision. The platforms are different enough that picking by cloud fit is usually right.