Best practices and features checklist for Teams agents

This article provides a list of best practices and features that developers should consider when implementing Teams agents.

These practices and features represent standard user expectations for agents in Teams and are broadly applicable. Agents that omit any of them should have a strong reason for doing so. Review each item to ensure your agent provides a great experience in Teams.

Message streaming and thinking indicators

Like many LLM chat experiences, Teams agents can stream messages in one-on-one conversations to maximize responsiveness. Before streaming begins, Teams can also display a customizable, updatable thinking indicator to indicate liveness and provide insight into how the message is being generated.

Targeted messaging

Targeted messaging enables agents to privately interact with users in group conversations.

An agent's ability to receive targeted messages must be enabled in its app manifest configuration, but should be considered a standard feature of Teams agents. Users will expect agents in group conversations to use targeted messaging to enable context-sensitive actions while maintaining privacy and keeping conversations focused. Agents should always consider the visibility of messages they receive and use it to inform response generation, response visibility, and context tracking behavior.

Emoji reactions

Reactions are lightweight emoji markers that all participants in a conversation, including agents, can attach to chat messages. Agents can use reactions to communicate acknowledgement or status without creating additional noise in a conversation.

Prompt starters

Prompt starters are prewritten prompts that help users understand an agent's capabilities and quickly begin interacting with it. Prompt starters are conveniently accessible and discoverable in all chat scenarios. In one-on-one chats, they are displayed prominently before the conversation starts for agents that do not send a welcome message.

Welcome messages

Agents can proactively send welcome messages when users add them to a conversation or install them as a personal app with one-on-one chat capabilities. Agents should use welcome messages to explain the value they provide, present configuration instructions, offer example prompts, and describe any commands they expose.

Agents distributed through the Teams Store and enabled for personal scope (one-on-one conversations) must implement welcome messages, prompt starters, or both.

Agent slash commands

Agents that provide command-style functionality can make it discoverable and easily usable with agent slash commands. An agent's commands are discoverable and invokable through Teams' slash command and @mention autocomplete menus in chat.

Suggested actions

Suggested actions are buttons on agent messages that enable users to quickly respond or take action. Unlike prompt starters and command names, agents can dynamically generate and attach suggested actions with each message they send, guiding users through workflows with context-aware options or intelligently suggesting next steps.

Suggested actions are key to enabling efficient and satisfying human-in-the-loop workflows, where an agent requests user approval or confirmation before taking action, but are useful in any scenario that benefits from quick user interaction. They can be configured to send a command-style chat response to the agent, silently invoke an agent behavior without a chat message, or insert a pre-constructed message into the user's compose box.

AI content labels

Teams' indicator for AI-generated messages should always be used when applicable. Correct usage of the AI indicator is a policy requirement for agents distributed through the Teams Store.

Sensitivity labels indicate messages that might contain confidential information, or links to resources only accessible to certain individuals.

Citations provide a standard way to surface and link any data sources used to generate a message.

User feedback controls

Agents can attach standardized feedback controls to their messages. Users can like or dislike messages and optionally provide detailed feedback.