Ranked list · How to choose · Pricing comparison · Methodology

Amazon Bedrock AgentCore vs Microsoft Foundry Agent Service (2026)

Answer in brief: Both are code-first platforms for running production agents built with any major framework. Pick Amazon Bedrock AgentCore if you are on AWS and want modular, per-second-billed building blocks: Runtime, Gateway, Identity, Memory, Policy, Evaluations and Agent Registry, each with a published price. Pick Microsoft Foundry Agent Service if you are on Azure and want managed prompt agents plus hosted code agents with Entra identities, which you can publish straight into Teams and Microsoft Copilot. AgentCore is #4 on our ranked shortlist. Foundry ships with Copilot Studio as our #2.

Top pick for AWS engineering teams: Amazon Bedrock AgentCore. No minimums, per-second billing, up to $200 in Free Tier credits for new AWS customers, and #3 overall in ISG's 2026 AI Agents Buyers Guide. Visit Amazon Bedrock AgentCore →
Top pick for Azure and Microsoft 365 teams: Microsoft Foundry Agent Service. Spans no-code prompt agents to fully containerized hosted agents, with an Entra identity for each agent. Visit Microsoft Foundry Agent Service →

This is independent research. It draws on AWS and Microsoft documentation and pricing pages checked on 2026-11-09. We have not tested these products hands-on.

Component by component

Capability	Amazon Bedrock AgentCore	Microsoft Foundry Agent Service
Frameworks	LangChain, OpenAI Agents SDK, Claude Agent SDK, Strands or your own	Agent Framework, LangGraph, OpenAI Agents SDK, Anthropic Agent SDK, Copilot SDK or custom code
Agent types	Runtime on serverless microVMs, or on Instances (EC2 plus a management fee)	Prompt agents, voice-based prompt agents (preview), hosted agents
Tools	Gateway turns APIs and Lambda into MCP tools; Browser, Code Interpreter, Web Search	Toolboxes shared through a managed MCP endpoint; web search, file search, code interpreter
Identity	AgentCore Identity; free when used through Runtime or Gateway	A dedicated Microsoft Entra identity per agent, plus RBAC
Policy	Cedar-based Policy checks every Gateway tool call	Content filters, guardrails, virtual network isolation
Quality	Evaluations with 13 built-in evaluators; Optimization with A/B tests	Tracing, evaluations, Application Insights; agent optimizer (preview)
Catalog	AWS Agent Registry	Publishes to the Entra Agent Registry, Teams and Microsoft Copilot
Cost model	A published price for each component	Per-call inference and tool usage, plus container compute for hosted agents

AgentCore: pay only while the agent is actually working

AgentCore's Runtime bills CPU only during active processing. CPU scales to zero during I/O waits, which AWS says take up 30–70% of a typical agent session's time. Published rates: Runtime v1 costs $0.0895 per vCPU-hour and $0.00945 per GB-hour. Runtime v2 costs $0.1276 per vCPU-hour and $0.0169 per GB-hour on consumption, with lower committed-baseline rates. Gateway costs $0.005 per 1,000 invocations, Web Search $7 per 1,000 queries, and Web Answer $10 per 1,000 tool calls. Policy costs $0.000025 per authorization request, and long-term memory retrieval $0.50 per 1,000. AWS's own worked example, 1 million ten-minute support sessions a month on Runtime v2, comes to $6,703 in runtime compute before model costs. Agent Registry's free tier covers the first 5,000 records each month.

Foundry: from prompt to container under one Entra identity

Foundry has three paths. You can configure a prompt agent with no code or infrastructure, run real-time voice agents through Voice Live, or ship your own code as a container or .zip that Foundry hosts with autoscaling. Agents can swap models from the Foundry catalog without code changes. Shared toolboxes give you one governed MCP endpoint, and finished agents publish to Teams and Microsoft Copilot. That last step is the bridge to Copilot Studio users. Microsoft was named a Leader in Gartner's first AI Agent Management Platforms Magic Quadrant.

Cost watch-outs

AgentCore: model inference, CloudWatch observability, ECR or S3 storage and data transfer are billed separately. Optimization Insights is free in preview, and pricing will be announced before general availability.
Foundry: costs combine model inference, tools and, for hosted agents, container compute. Voice and some evaluation features are in preview without an SLA.

Choose AgentCore if…

Your data, IAM and compliance programs are on AWS.
You want to adopt components one at a time, such as Gateway and Identity only.
You need cost-per-session modeling from a public rate card.

Choose Foundry if…

Entra ID governs your identities and Azure hosts your data.
You want agents to reach employees in Teams and Microsoft Copilot.
You want a no-code prompt-agent tier and pro-code hosting side by side.

Comparing raw model hosting rather than agent runtimes? See our sibling guide to the best inference providers.

FAQ

Do both support MCP?

Yes. AgentCore Gateway lets agents securely call MCP servers, APIs and Lambda functions. Foundry toolboxes share tools, including MCP servers, through one managed MCP endpoint.

Is there a free way to start?

New AWS customers get up to $200 in Free Tier credits, and AgentCore has no minimum fees. Foundry prompt agents have no compute to manage; you pay for inference and tools.

Can I use Anthropic or OpenAI frameworks on either?

Yes. AWS lists the OpenAI Agents SDK and Claude Agent SDK. Microsoft lists the OpenAI Agents SDK and Anthropic Agent SDK for hosted agents.

Keep exploring

Visit Amazon Bedrock AgentCore → · Visit Microsoft Foundry Agent Service →

Sources

Last updated: 2026-11-09.