Microsoft Agent Stack — Selection and Interop
The five pieces solve different layers of the same problem. Most production systems combine
a build layer (Copilot Studio or Agent Framework) with a hosting/channel layer
(Foundry Agent Service or M365 Agents SDK/Teams SDK).
Decision matrix
Source: https://learn.microsoft.com/microsoft-365/copilot/extensibility/overview-custom-engine-agent
|
Copilot Studio |
Teams SDK |
M365 Agents SDK |
Foundry |
| Approach |
Low-code SaaS |
Pro-code |
Pro-code |
Low-code or pro-code PaaS |
| Orchestrator |
Copilot Studio |
Bring your own |
Bring your own (Agent Framework, SK, LangChain) |
Bring your own / managed |
| Channels |
M365 Copilot, Teams, web, custom (Direct Line) |
M365 Copilot, Teams |
M365 Copilot, Teams + 10+ channels via Azure Bot Service |
M365 Copilot, Teams (others via custom integration) |
| Languages |
n/a |
C#, TS/JS, Python (preview) |
C#, JS, Python |
Python, C# |
Rules of thumb
- Maker-built, governed, Power-Platform-integrated, department scale → Copilot Studio
(Copilot Credits abstract LLM token pricing).
- Code-level control of the agent loop, multi-agent workflows, provider flexibility →
Agent Framework (it's the library, not the host).
- Azure-managed runtime, model catalog, enterprise identity/observability → Foundry
Agent Service (CAF decision tree: SaaS → Copilot Studio; PaaS pro-code → Foundry).
- Same agent on many channels (the Bot Framework succession path) → M365 Agents SDK.
- Deeply Teams-native collaborative experience → Teams SDK.
Interop patterns
| Pattern |
How |
| Agent Framework inside M365 Agents SDK |
AgentApplication turn handler calls agent.run(...); SDK owns channels/auth, framework owns reasoning |
| Agent Framework on Foundry |
Hosted agent via agent-framework-foundry-hosting (ResponsesHostServer) — managed endpoint, Entra agent identity |
| Foundry agent in M365/Teams |
Publish from the Foundry portal (auto-provisions Azure Bot Service + Entra) or an Agents Toolkit proxy app |
| Copilot Studio → Foundry |
Connect an external Foundry agent as a child agent (preview; new-portal agents only) |
| Code → Copilot Studio |
CopilotStudioAgent (agent-framework-copilotstudio); M365 Agents SDK OBO-auth samples |
| Humans in the loop from any stack |
Teams as an MCP server: notify / ask / request_approval |
The two protocols
- MCP is the common tool fabric — Copilot Studio adds MCP servers as tools, Foundry has
an MCP tool type, Agent Framework ships MCP clients (
MCPStreamableHTTPTool et al.), and
Teams itself can be exposed as an MCP server. Build a capability once as an MCP server
and every stack can consume it.
- A2A is the common agent-to-agent protocol — Agent Framework
A2AAgent /
Microsoft.Agents.AI.A2A, Foundry's A2A tool (preview), Teams SDK bot-to-bot A2A.
Use A2A when two agents converse; use MCP when an agent calls a capability.
Cross-cutting: Agent 365
The Agent 365 SDK (@microsoft/agents-a365-*) provides Entra-based agent identity, an
agent registry, and OpenTelemetry observability for agents from any framework —
including agents registered from Google Vertex AI or Amazon Bedrock. Reach for it when an
organization needs one governance plane over heterogeneous agents.
Architecture review checklist
When reviewing a proposed multi-stack design:
- Exactly one owner per layer — conversation/channel, orchestration, tools, hosting,
identity. Two components owning the same layer is the #1 design smell.
- Every cross-stack hop uses MCP or A2A (or a documented connector) — no bespoke HTTP
bridges that bypass identity and telemetry.
- Identity flows end-to-end: user → channel → agent (Entra agent identity) → downstream
scopes, with on-behalf-of where user context must survive.
- One telemetry plane (OpenTelemetry / Agent 365) across all stacks in the design.
- Cost model named per layer: Copilot Credits, Foundry consumption, model tokens.
1---2name: agent-interop3description: Choose and combine Microsoft agent stacks — the Copilot Studio vs Teams SDK vs M365 Agents SDK vs Agent Framework vs Foundry decision matrix, plus interop patterns - MCP as the tool fabric, A2A as the agent protocol, hosting matrices, and Agent 365 identity/observability.4---56# Microsoft Agent Stack — Selection and Interop78The five pieces solve different layers of the same problem. Most production systems combine9a **build layer** (Copilot Studio or Agent Framework) with a **hosting/channel layer**10(Foundry Agent Service or M365 Agents SDK/Teams SDK).1112## Decision matrix1314Source: https://learn.microsoft.com/microsoft-365/copilot/extensibility/overview-custom-engine-agent1516| | Copilot Studio | Teams SDK | M365 Agents SDK | Foundry |17|---|---|---|---|---|18| Approach | Low-code SaaS | Pro-code | Pro-code | Low-code or pro-code PaaS |19| Orchestrator | Copilot Studio | Bring your own | Bring your own (Agent Framework, SK, LangChain) | Bring your own / managed |20| Channels | M365 Copilot, Teams, web, custom (Direct Line) | M365 Copilot, Teams | M365 Copilot, Teams + 10+ channels via Azure Bot Service | M365 Copilot, Teams (others via custom integration) |21| Languages | n/a | C#, TS/JS, Python (preview) | C#, JS, Python | Python, C# |2223**Rules of thumb**2425- Maker-built, governed, Power-Platform-integrated, department scale → **Copilot Studio**26 (Copilot Credits abstract LLM token pricing).27- Code-level control of the agent loop, multi-agent workflows, provider flexibility →28 **Agent Framework** (it's the library, not the host).29- Azure-managed runtime, model catalog, enterprise identity/observability → **Foundry30 Agent Service** (CAF decision tree: SaaS → Copilot Studio; PaaS pro-code → Foundry).31- Same agent on many channels (the Bot Framework succession path) → **M365 Agents SDK**.32- Deeply Teams-native collaborative experience → **Teams SDK**.3334## Interop patterns3536| Pattern | How |37|---|---|38| Agent Framework inside M365 Agents SDK | AgentApplication turn handler calls `agent.run(...)`; SDK owns channels/auth, framework owns reasoning |39| Agent Framework on Foundry | Hosted agent via `agent-framework-foundry-hosting` (`ResponsesHostServer`) — managed endpoint, Entra agent identity |40| Foundry agent in M365/Teams | Publish from the Foundry portal (auto-provisions Azure Bot Service + Entra) or an Agents Toolkit proxy app |41| Copilot Studio → Foundry | Connect an external Foundry agent as a **child agent** (preview; new-portal agents only) |42| Code → Copilot Studio | `CopilotStudioAgent` (`agent-framework-copilotstudio`); M365 Agents SDK OBO-auth samples |43| Humans in the loop from any stack | Teams as an MCP server: `notify` / `ask` / `request_approval` |4445## The two protocols4647- **MCP is the common tool fabric** — Copilot Studio adds MCP servers as tools, Foundry has48 an MCP tool type, Agent Framework ships MCP clients (`MCPStreamableHTTPTool` et al.), and49 Teams itself can be exposed as an MCP server. Build a capability once as an MCP server50 and every stack can consume it.51- **A2A is the common agent-to-agent protocol** — Agent Framework `A2AAgent` /52 `Microsoft.Agents.AI.A2A`, Foundry's A2A tool (preview), Teams SDK bot-to-bot A2A.53 Use A2A when two *agents* converse; use MCP when an agent calls a *capability*.5455## Cross-cutting: Agent 3655657The Agent 365 SDK (`@microsoft/agents-a365-*`) provides Entra-based **agent identity**, an58agent **registry**, and **OpenTelemetry observability** for agents from any framework —59including agents registered from Google Vertex AI or Amazon Bedrock. Reach for it when an60organization needs one governance plane over heterogeneous agents.6162## Architecture review checklist6364When reviewing a proposed multi-stack design:65661. Exactly one owner per layer — conversation/channel, orchestration, tools, hosting,67 identity. Two components owning the same layer is the #1 design smell.682. Every cross-stack hop uses MCP or A2A (or a documented connector) — no bespoke HTTP69 bridges that bypass identity and telemetry.703. Identity flows end-to-end: user → channel → agent (Entra agent identity) → downstream71 scopes, with on-behalf-of where user context must survive.724. One telemetry plane (OpenTelemetry / Agent 365) across all stacks in the design.735. Cost model named per layer: Copilot Credits, Foundry consumption, model tokens.
markus41