AgentCore vs Foundry Agent Service: what the runtime choice turns on

Updated 19 November 2026 · Editorial assessment

At a glance

AgentCore vs Foundry Agent Service: what the runtime choice turns on
AttributeAmazon Bedrock AgentCoreMicrosoft Foundry Agent Service
What it isAgentCore is an agentic platform for building, deploying and operating agents securely at scale using any framework and foundation model.Foundry Agent Service enables professional developers to build, deploy and scale AI agents natively in Foundry.
Frameworks and languagesAgentCore Runtime works with custom frameworks and any open-source framework, including CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI Agents SDK and Strands Agents.Microsoft's hosted agent documentation lists Python and C# as the supported languages. The Foundry overview says hosted agents can be built with Agent Framework, LangGraph, the OpenAI Agents SDK, the Anthropic Agent SDK or your own code.
Session modelAgentCore gives each session its own dedicated microVM, and sessions last up to 8 hours for each lifecycle on microVMs, or up to 14 days on Instances.According to Microsoft, hosted agents scale per session, and the platform creates a VM-isolated sandbox for each session on demand. Microsoft documents an idle timeout that defaults to 15 minutes and can be set from 2 through 240 minutes.
MemoryAgentCore Memory supports short-term and long-term memory, and memory stores can be shared across agents.Microsoft describes memory in Foundry Agent Service as a managed, long-term memory solution, and both it and the Memory Store API are in preview.
ProtocolsAgentCore Runtime works with protocols like MCP and A2A, and for A2A it acts as a proxy between clients and your A2A server.Microsoft's A2A guide recommends the generally available a2a tool type for new integrations, and its MCP guide says Agent Service supports public and private MCP server endpoints.
IdentityAgentCore Identity is compatible with existing identity providers, and its page names Amazon Cognito, Okta, Microsoft Azure Entra ID and Auth0 as examples.Microsoft says Agent Service assigns a hosted agent a dedicated Microsoft Entra ID when you deploy it.
Tool accessAgentCore Gateway is a fully managed AI gateway that gives agents one secure entry point to tools, other agents and models.Microsoft describes Toolbox in Foundry as a way to define a curated set of tools once and expose them through a single MCP-compatible endpoint.
ObservabilityAgentCore emits telemetry in OpenTelemetry-compatible format, and its metrics, spans and logs are stored in Amazon CloudWatch.Hosted agents that use the protocol libraries emit OpenTelemetry traces by default, and those traces appear in the linked Application Insights resource.
BillingAgentCore offers consumption-based pricing with no upfront commitments or minimum fees. Runtime offers two compute types, consumption-based microVMs and Instances billed at EC2 instance cost plus a management fee.Foundry Agent Service bills hosted agents on container compute consumed per hour, and model token consumption is charged separately. Per Microsoft's documentation, hosted agent billing is based on CPU + memory consumed across all active sessions.

How does each one coordinate agents?

AgentCore Runtime lists support for multi-modal and multi-agent workloads and works with custom and open-source frameworks. If you prefer a managed loop, Harness lets you define an agent with a model, system prompt and tools, and it handles orchestration, tool execution and memory management. Agents can also share memory stores through AgentCore Memory.

AgentCore passes JSON-RPC payloads from the InvokeAgentRuntime API to the A2A container without modification. Gateway fronts other agents and HTTP services through passthrough targets, including A2A traffic.

Foundry Agent Service coordinates through tools and workflows. The A2A tool lets a Foundry agent exchange context and collaborate with external agents, and the Foundry agent uses the remote agent's response to answer the user and continues to manage the conversation. Workflows orchestrate multiple Foundry agents declaratively using sequential, group chat or human-in-the-loop patterns. For custom logic, the overview names multi-agent systems as a secondary use of hosted agents, where Foundry handles hosting, scaling and identity.

When should you choose AgentCore?

When should you choose Foundry Agent Service?

Can you use them together?

Both sit in the runtime layer of this site's category map, so a shared system would run some agents on each and connect them over a protocol both support. On the AWS side, an AgentCore A2A server supports both SigV4 and OAuth 2.0. That makes an AgentCore A2A server a candidate target for a Foundry agent, but test authentication and protocol version before you design around it.

Tools are a second meeting point. A Foundry Toolbox endpoint is MCP-compatible, so any runtime that can consume an MCP server can also consume a Toolbox. Microsoft's MCP guide says agents authenticate to the Toolbox endpoint itself by using Microsoft Entra credentials, such as DefaultAzureCredential. Plan that credential path for any agent outside Azure, and see the guide to connecting agents across clouds for the wider pattern.

What should you watch out for?

What do enterprise buyers ask?

Can AgentCore host an A2A server?

Yes. AgentCore Runtime lets you deploy and run Agent-to-Agent servers, and it expects them on port 9000 at the root path.

Does Foundry Agent Service charge to create a prompt agent?

Its pricing page states no additional charge for creating or running Foundry-native agents using prompts and workflows. Model tokens are charged, and Foundry Tools and Foundry IQ connections carry separate charges and licenses.

Is AgentCore Identity billed separately?

Not when you use it with Runtime or Gateway, because AgentCore Identity is included at no additional cost in that case.

Where do I compare these with other layers?

Use the comparison table for the full set of runtimes, gateways and event meshes, and the RFP checklist to turn these points into requirements.

Which sources does this page cite?

Reviewed Nov 2026