Microsoft Foundry Agent Service
Foundry Agent Service is the managed runtime in Microsoft Foundry for building, hosting and scaling AI agents. Teams configure prompt agents with no code, run voice agents, or deploy their own agent code as hosted containers, with Entra identity, tools, tracing and publishing to Teams and Microsoft Copilot.1
Also known as Foundry Agent Service, Azure AI Agent Service
- Intended for
- Agent and application developers, Platform engineers, Solution architects, Security and identity teams
- Runs on
- Azure public cloud regions (hosted agents only in the regions Learn lists), Private networking: standard setup for prompt agents, bring-your-own VNet for hosted agents2
At a glance
- What is it?
- Microsoft Foundry Agent Service is a managed platform inside Microsoft Foundry for building, deploying and scaling AI agents. It offers three agent types: prompt agents defined by instructions, a model and tools; voice-based prompt agents that hold real-time spoken conversations through Voice Live; and hosted agents, where you package your own code built with Microsoft Agent Framework, LangGraph, the OpenAI Agents SDK, the Anthropic Agent SDK, the GitHub Copilot SDK or custom code, and Foundry runs it. The current service uses the Responses API (Agents v2), which replaced the Assistants API. Earlier SDKs used the name Azure AI Agent Service, and the older agents now run as Foundry Agent Service (classic).134
- What does it do?
- The service hosts and scales agents, manages conversations, tool calls and versions, and can give each agent its own Microsoft Entra identity. Toolboxes group tools such as web search, file search, code interpreter, MCP servers and custom functions behind one managed MCP endpoint. Tracing, evaluations and Application Insights show what each agent did. Published agents get a stable endpoint and can be shared to Microsoft Teams, Microsoft Copilot and the Entra Agent Registry, and agents can talk to other agents over the A2A protocol.1
- Who needs it?
- Developers and platform teams who build custom agents that need a choice of model, integration with company systems or custom logic, and who prefer a managed runtime to running their own containers or Kubernetes. Microsoft's Cloud Adoption Framework places Foundry at the strategic transformation end, with deep integration and custom logic, and Copilot Studio at the process transformation end with low-code tools.5
- What does it need?16
- An Azure subscription with a Foundry resource and project
- Microsoft Entra ID authentication, which agents require instead of API keys
- A supported model from the Foundry catalog
- For hosted agents: a container image or source .zip in Python or C#, and a region that offers hosted agents
- For private networking: the standard setup for prompt agents or a bring-your-own virtual network for hosted agents
- What it is not
- Agent Service is not a low-code maker tool for business users; Microsoft positions Copilot Studio for that. It is not an agent framework: you write orchestration with Microsoft Agent Framework or another SDK and host it here. It is not the place for new multi-agent workflows: Microsoft is retiring Foundry workflows on 1 December 2026 and points new development to Microsoft Agent Framework. And it is not always needed: Microsoft's guidance says to check first whether a ready-made SaaS agent meets the requirement.56
Availability and licensing. Agents (core) and publishing to Microsoft Copilot and Teams are GA; Voice Live agents, memory and agent guardrails are preview; the agent optimizer is limited preview (Learn GA overview, checked 26 Oct 2026). The hosted agents page has no preview label. Workflows retire on 1 December 2026; agents (classic) retire on 31 March 2027. Pricing: inference and tools, plus CPU and memory for hosted agents.267
The problem it solves
Teams that build agents with open-source frameworks also have to handle containers, web servers, security, memory, scaling, instrumentation and version rollback. Microsoft's hosted agent documentation names these cross-cutting jobs as the problem the service takes on.
Agent Service handles the hosting side. Simple agents need no runtime code at all; for code-based agents the platform provides the endpoint, per-session isolation, identity, state and observability while your code keeps control of the logic.2
How it works
- Prompt agents: defined by configuration in the portal, SDK or REST; Foundry runs them with no runtime code. Cost: per-call inference plus tool usage.
- Voice-based prompt agents (preview): connect through Voice Live over a real-time WebSocket, using native speech-to-speech or a cascaded speech, text and speech design; phone numbers through Teams Phone extensibility or Twilio.
- Hosted agents: you push a container image to Azure Container Registry, or a .zip that Foundry builds. Each session runs in its own VM-isolated sandbox with a persistent $HOME, sized from 0.5 vCPU and 1 GiB up to 2 vCPU and 4 GiB, in Python or C#. Compute is released after an idle timeout (default 15 minutes) and state is restored on resume. Cost adds CPU and memory during active sessions.
Endpoints support the Responses, Invocations, Activity and A2A protocols. Prompt agents can use private networking; hosted agents use a bring-your-own virtual network.12
Diagram as a list
Copilot & business apps
- Apps, Microsoft Teams and Microsoft CopilotWhere users and other systems reach the agentConnects to Agent endpoint (Responses, Invocations, Activity, A2A)
Agents & automation
- Agent endpoint (Responses, Invocations, Activity, A2A)Stable entry point for a published agent versionConnects to Agent runtime (prompt and hosted agents)
- Agent runtime (prompt and hosted agents)Hosts agents, manages conversations, tool calls and sessions
AI models & platform
- Foundry ModelsModels used for reasoning
- Toolbox MCP endpointCurated tools such as web search, code interpreter and MCP servers
Data & analytics
- Foundry IQ knowledge basesGrounding in company data through Azure AI Search
Developer tools
- Tracing and Application InsightsTraces, metrics and evaluations
Security, identity & governance
- Microsoft Entra agent identityScoped, auditable access to resources
Capabilities
Prompt agents1
Agents defined entirely by instructions, model choice and tools, authored in the portal or as code in a CI/CD pipeline.
Why it matters: The fastest path to a production agent with nothing to host.
Limits: No custom orchestration logic; use a hosted agent for that.
Hosted agents2
Your own agent code, in any framework, runs in per-session VM-isolated sandboxes with a dedicated endpoint, Entra identity and persisted session state.
Why it matters: Full control of logic while Foundry handles hosting, scaling and identity.
Limits: Python and C# only; sandboxes go up to 2 vCPU and 4 GiB; an endpoint serves one version with no traffic splitting.
Voice-based prompt agents16
Managed real-time voice agents through Voice Live, with browser testing and phone channels via Teams Phone extensibility or Twilio.
Why it matters: Customer-facing voice experiences without hosting voice orchestration.
Limits: Preview; direct SIP is not offered, and phone callers receive audio only.
Toolboxes and MCP16
Curated tools shared across agents through one managed MCP endpoint with central authentication, governance and versioning.
Why it matters: Change a tool once instead of rewiring every agent.
Limits: Authentication options are key, Entra identity, OAuth passthrough or unauthenticated; non-Microsoft tools may send data outside your boundary.
Identity and networking12
Dedicated Microsoft Entra identities for agents, RBAC, content filters and virtual network isolation.
Why it matters: Agents get scoped access to resources without shared credentials.
Limits: A private Azure Container Registry for hosted agents works only for projects created after 25 June 2026.
Publishing and protocols16
Versioned agents with stable endpoints, shared to Microsoft Teams, Microsoft Copilot and the Entra Agent Registry; A2A for agent-to-agent calls.
Why it matters: Agents reach users where they already work and can be called by other agents.
Limits: A2A v1.0 is GA but v0.3 is preview; voice-based agents do not show the standard Teams publishing card.
Observability and optimization16
End-to-end tracing, evaluations and Application Insights integration, plus an agent optimizer that proposes better instructions and tool descriptions.
Why it matters: Teams can see each decision and measure quality over time.
Limits: Tracing is GA for prompt and hosted agents; the agent optimizer is limited preview.
Memory67
Managed short-term, long-term and procedural memory for agents.
Why it matters: Agents can keep useful context across conversations.
Limits: Memory is preview (checked 26 Oct 2026).
Practical use cases
A contact center wants an automated voice agent for routine calls without building speech infrastructure.
- Approach
- Create a voice-based prompt agent and connect a phone number through Teams Phone extensibility or Twilio.
- Role of Microsoft Foundry Agent Service
- Agent Service manages the agent and voice orchestration through Voice Live.
- Data, infrastructure and skills
- Foundry project, Voice Live model, Application Insights and a telephony channel.
- Type of benefit
- Faster customer response
- Caveats
- Voice agents are preview; direct SIP is not supported.
- First step
- Build a voice agent and test it from the browser playground.
Sources 1
Employees need an agent in Teams that answers questions and performs simple tasks in internal systems.
- Approach
- Build a prompt agent with toolbox tools and a knowledge base, then publish it to Teams.
- Role of Microsoft Foundry Agent Service
- Agent Service hosts the agent, applies Entra identity and handles publishing.
- Data, infrastructure and skills
- Foundry project, model deployment, tools and Teams publishing rights.
- Type of benefit
- Employee time saved
- Caveats
- Agents require Microsoft Entra ID authentication; review data flows to any non-Microsoft tools.
- First step
- Prototype the agent in the playground with two or three tools.
External systems such as GitHub or Jira should trigger an agent with their own payloads.
- Approach
- Deploy a hosted agent that exposes the Invocations protocol to receive webhooks.
- Role of Microsoft Foundry Agent Service
- Agent Service runs the container per session and provides identity and tracing.
- Data, infrastructure and skills
- Agent code in Python or C#, Azure Container Registry and a supported region.
- Type of benefit
- Process automation
- Caveats
- You maintain the agent code; billing includes CPU and memory per active session.
- First step
- Deploy a sample hosted agent from the foundry-samples repository.
Sources 2
Works with
Requires
- Microsoft Foundry ModelsAgents use supported models from the Foundry model catalog.
Optional integration
- Microsoft Agent FrameworkHosted agents can run Agent Framework code, and Microsoft directs workflow users to Agent Framework.
- Azure AI SearchAgents ground on Foundry IQ knowledge bases and can bring their own Azure AI Search resource.
- Azure Cosmos DBBring-your-own resources include Azure Cosmos DB for conversation state.
- Microsoft TeamsAgents can be published to Microsoft Teams.
- Microsoft 365 CopilotAgents can be published to Microsoft Copilot (the renamed Microsoft 365 Copilot app).
- Microsoft EntraEach agent can have a dedicated Microsoft Entra identity and appear in the Entra Agent Registry.
Complementary tools
- Microsoft Agent 365Foundry agents are among the Microsoft platforms Agent 365 integrates with.
Same family
- Microsoft FoundryThe product page lists Foundry Agent Service as a service built into Foundry.
Alternative approaches
- Microsoft Copilot StudioMicrosoft presents Copilot Studio (low-code SaaS) and Foundry (PaaS) as alternative build paths.
Optional integration for
- Azure AI SearchFoundry IQ knowledge bases built on Azure AI Search connect to Agent Service agents.
- Microsoft Foundry ModelsAgent Service works with models from the Foundry catalog and lets you swap models without code changes.
- Microsoft Agent FrameworkAgent Framework code can run as a Foundry hosted agent.
- Azure Cosmos DBStandard agent setup stores threads in a customer-owned Cosmos DB for NoSQL account.
- Foundry ToolsVoice-based agents use Azure Speech in Foundry Tools when the architecture needs transcription or synthesis.
Relationship labels follow Microsoft's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Compared with
Getting started
Set up the environment
Create a Foundry project and choose the basic setup for prototypes or the standard setup for enterprise data and networking control.
Check: You can open the project in the Foundry portal.
Create a prompt agent
Use the portal or the prompt agent quickstart to set instructions, a model and tools.
Check: The agent replies correctly in the agents playground.
Add a toolbox
Curate the tools the agent needs into a toolbox and attach it.
Check: Traces show the expected tool calls.
Move custom logic to a hosted agent
Package your framework code as a container or .zip and deploy it with the Azure Developer CLI.
Check: The agent's dedicated Responses endpoint answers test requests.
Publish
Promote a version to a managed resource and publish it to Microsoft Teams and Microsoft Copilot.
Check: Users can find and chat with the agent in Teams.
Leave workflows before 1 December 2026
If you use Foundry workflows, follow the migration guide to Microsoft Agent Framework.
Check: No production process depends on Foundry workflows.
Official resources
Deciding whether you need it?
Three free tools apply Microsoft's own published guidance to your situation, and say when Microsoft Foundry Agent Service or any new technology is not needed.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
- What is Microsoft Foundry Agent Service? (opens in a new tab)
- Hosted agents in Foundry Agent Service (opens in a new tab)
- What is Microsoft Foundry? (opens in a new tab)
- What's new in Foundry Agent Service? (classic) (opens in a new tab)
- Technology plan for AI agents (Cloud Adoption Framework) (opens in a new tab)
- Microsoft Foundry portal general availability overview (opens in a new tab)
- Foundry Agent Service product page (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.