Lobbi-Docs Teams Agents — Teams SDK
The Teams AI Library is now renamed to Teams SDK (branded "Teams AI Library v2" in some docs) — rebuilt from the ground up for Teams-native agents. GA for JavaScript and C#; Python is developer preview. v1 is deprecated. Docs: https://learn.microsoft.com/microsoftteams/platform/teams-ai-library/welcome/overview · Repos: microsoft/teams.ts, microsoft/teams.py
Packages and CLI
- npm:
@microsoft/teams.apps(theAppclass),@microsoft/teams.api(MessageActivityand friends),@microsoft/teams.common - CLI:
npm i -g @microsoft/teams.cli@preview→teams new my-agent --template echo
Bring your own AI. The SDK deprecated its own AI packages (@microsoft/teams.ai
ChatPrompt/Model, old @microsoft/teams.mcp, @microsoft/teams.a2a). The documented
pattern: run the agent loop with the OpenAI SDK or Microsoft Agent Framework
(skills/agent-framework) and wire MCP/A2A directly — Teams SDK handles activity routing
and Teams affordances.
Minimal agent
import { App } from '@microsoft/teams.apps';
const app = new App();
app.on('message', async ({ send, activity }) => {
await send({ type: 'typing' });
await send(`you said "${activity.text}"`);
});
app.start(process.env.PORT || 3978).catch(console.error);
Agent Framework loop streaming into Teams (Python preview)
@app.on_message
async def handle_message(ctx: ActivityContext[MessageActivity]):
async for chunk in agent.run(ctx.activity.text or "", stream=True):
if chunk.text:
ctx.stream.emit(chunk.text)
ctx.stream.emit(MessageActivityInput().add_ai_generated().add_feedback(mode="custom"))
Teams-specific capabilities
| Capability | Notes |
|---|---|
| Adaptive Cards | Dialogs, message extensions, link unfurling — validate card JSON against the schema |
| Streaming | ctx.stream.emit(...) for progressive responses |
| AI-generated label | .addAiGenerated() — required UX for AI output in Teams |
| Feedback controls | .addFeedback('custom') — thumbs/feedback affordances on agent messages |
| Citations | Attach sources to generated answers |
| Prompt starters | Suggested first prompts in the agent's Teams surface |
| Proactive messages | Agent-initiated notifications to users/channels |
| Meetings | Meeting-aware agents (join context, transcripts where permitted) |
| Teams as MCP server | Expose Teams to agents: notify / ask / request_approval reach humans |
| Bot-to-bot A2A | Agent-to-agent conversations across Teams bots |
Relationship to the M365 Agents SDK
Teams SDK is Teams-centric (collaborative, multi-user, meetings, real-time). To extend the
same agent to Outlook, M365 Copilot, and other channels, use the M365 Agents SDK
(skills/m365-agents-sdk) — which also ships microsoft-agents-hosting-teams for the
Teams surface. Both build and publish through the Microsoft 365 Agents Toolkit.
Rule of thumb: Teams-only experience → Teams SDK; multi-channel custom engine agent → M365 Agents SDK, adding Teams affordances through its Teams hosting package.
Practices
- Always attach the AI-generated label and feedback controls to model-generated messages — reviewers and store validation look for both.
- Stream long responses; Teams users perceive non-streaming agents as hung.
- Keep card payloads small and versioned; test dark mode and mobile rendering.
- Migrating from Teams AI v1 or Teams Toolkit projects: the toolkit is now the
Microsoft 365 Agents Toolkit — see
/msagent-migratefor the project-structure map.