M365 Agents Sdk

Microsoft 365 Agents SDK

The channel and hosting layer for custom engine agents: receives and sends messages across Microsoft 365 Copilot, Teams, web, and 10+ third-party channels (Slack, Facebook Messenger, Twilio, SMS, email) via Azure Bot Service, which translates channel traffic into Activities. It is the successor to the Bot Framework SDK (BF SDK and Emulator are archived; support ended 2025-12-31). AI-agnostic: host Microsoft Agent Framework, Semantic Kernel, LangChain, OpenAI Agents, or Foundry agents inside it. Docs: https://learn.microsoft.com/microsoft-365/agents-sdk/agents-sdk-overview

Status (mid-2026): GA — .NET v1.6.x (.NET 8+), JavaScript v1.6.x (Node 18+), Python v1.1.x.

Packages

Core concepts

• Activity Protocol — the standard JSON message format shared across channels; typed models live in Microsoft.Agents.Core / @microsoft/agents-activity.
• AgentApplication — your entry point. Pipeline: Channel → hosting layer (HTTP auth) → AgentApplication routing → your handlers, with turn state loaded before and saved after each handler. Route ranking (RouteRank.Last) orders catch-alls.
• CloudAdapter processes /api/messages; agents are stateless with pluggable storage (memory, Blob, Cosmos).
• Tooling: Microsoft 365 Agents Toolkit (formerly Teams Toolkit — VS, VS Code, CLI) for scaffolding/manifests/publishing, and the Microsoft 365 Agents Playground — a local Teams-like sandbox needing no tenant, tunnel, or bot registration.
• Agent 365 SDK (@microsoft/agents-a365-*) layers on enterprise observability (OpenTelemetry), notifications, MCP tool-server management, and Entra-based agent identity.

Minimal agents

public class EchoAgent : AgentApplication
{
    public EchoAgent(AgentApplicationOptions options) : base(options)
        => OnActivity(ActivityTypes.Message, OnMessageAsync, rank: RouteRank.Last);

    private async Task OnMessageAsync(ITurnContext tc, ITurnState ts, CancellationToken ct)
        => await tc.SendActivityAsync($"You said: {tc.Activity.Text}", cancellationToken: ct);
}

import { AgentApplication, MemoryStorage, TurnContext, TurnState } from '@microsoft/agents-hosting'
import { startServer } from '@microsoft/agents-hosting-express'

const app = new AgentApplication<TurnState>({ storage: new MemoryStorage() })
app.onActivity('message', async (ctx: TurnContext) => ctx.sendActivity(`You said: ${ctx.activity.text}`))
startServer(app)

Hosting an AI engine inside

The SDK is plumbing; the intelligence is whatever you host. The documented pattern for Agent Framework / Semantic Kernel: instantiate the engine agent once, then call it from the message handler and relay the result as an activity. Guide: https://learn.microsoft.com/microsoft-365/agents-sdk/using-semantic-kernel-agent-framework

Keep the layers separate: channel/UX concerns (activities, cards, auth) in the AgentApplication handler; reasoning/tools in the engine (see skills/agent-framework).

Bot Framework migration

Coming from Bot Framework v4 (see /msagent-migrate):

• The Activity protocol carries over — message shapes and channel semantics are familiar.
• ActivityHandler/adapter wiring becomes AgentApplication routing; middleware moves to the new pipeline.
• Dialogs have no direct successor — replace waterfall dialogs with an LLM engine + tools, or explicit state machines for strictly deterministic flows.
• Re-register the messaging endpoint against Azure Bot Service; update auth to the current Microsoft.Agents.Authentication / MSAL configuration.

Practices

• Develop against the Agents Playground first; only stand up tunnels/bot registrations when channel-specific behavior needs testing.
• Store per-conversation state via the SDK's storage abstractions, not process memory, so scale-out works.
• One agent codebase, many channels: guard Teams-only affordances (cards, meetings) behind channel checks — see skills/teams-agents for the Teams-native layer.
• Publishing to M365 Copilot/Teams flows through app manifests and the Agents Toolkit — see /msagent-deploy.