AI Agents Overview
AI Agents enable intelligent, automated interactions within your application. They can process user messages, trigger tools, and respond with contextually relevant information. For a broader introduction, see the AI Agents section.Note: Currently, an Agent only responds to Text Messages.
1:1 and group conversations. AI Agents work in both one-on-one and group conversations. In a 1:1 chat, the end user talks directly with the agent user. In a group, the agent participates as a member — its messages (including cards) are delivered and attributed like any other member’s message.
Agent Run Lifecycle and Message Flow
This section explains how a user’s text message to an Agent becomes a structured “run” which emits real-time events and then produces agentic messages for historical retrieval.- A user sends a text message to an Agent.
- The platform starts a run and streams real-time events via the
AIAssistantListener. - After the run completes, persisted Agentic Messages arrive via the
MessageListener.
Real-time Events
Events are received via theonAIAssistantEventReceived method of the AIAssistantListener class in this general order:
- Run Start
- Zero or more tool call cycles (repeats for each tool invocation):
- Tool Call Start
- Tool Call Arguments
- Tool Call End
- Tool Call Result
- Zero or more card generation cycles (repeats for each card produced):
- Card Start
- Card (full payload)
- Card End
- One or more assistant reply streams:
- Text Message Start
- Text Message Content (multiple times; token/char streaming)
- Text Message End
- Run Finished
Run StartandRun Finishedare always emitted.Tool Callevents appear only when a backend or frontend tool is invoked. There can be multiple tool calls in a single run.Cardevents appear only when the agent produces a card. The UI can show a loading state onCard Start, render the card onCard(payload), and finalize onCard End.Text Messageevents are always emitted and carry the assistant’s reply incrementally.
- Dart
Event descriptions
- Run Start: A new run has begun for the user’s message.
- Tool Call Start: The agent decided to invoke a tool.
- Tool Call Arguments: Arguments being passed to the tool.
- Tool Call End: Tool execution completed.
- Tool Call Result: Tool’s output is available.
- Card Start: The agent started generating a card. Contains
cardIdandexecutionText(a human-readable status like “Building your product card…”). - Card: The full card payload is available. Use
getCard()to retrieve the raw card JSON and pass it to the renderer. - Card End: The card generation flow is finalized.
- Text Message Start: The agent started composing a reply.
- Text Message Content: Streaming content chunks for progressive rendering.
- Text Message End: The agent reply is complete.
- Run Finished: The run is finalized; persisted messages will follow.
Card Streaming Event Classes
| Class | Properties |
|---|---|
AIAssistantCardStartedEvent | runId, threadId, streamMessageId, cardId, executionText |
AIAssistantCardReceivedEvent | runId, threadId, streamMessageId, cardId, getCard() → Map<String, dynamic>? |
AIAssistantCardEndedEvent | runId, threadId, streamMessageId, cardId |
Agentic Messages
These events are received via theMessageListener after the run completes.
AIAssistantMessage: The full assistant reply.AIToolResultMessage: The final output of a tool call.AIToolArgumentMessage: The arguments that were passed to a tool.
- Dart
Starting from SDK version 5.0.5, AI assistant message content is delivered via
getElements(). If the agent response contains only a card (no accompanying text), getText() will return an empty string — always prefer getElements() as the primary data source for rendering.AIAssistantMessage Elements
A persistedAIAssistantMessage carries its content in two ways:
getText()— The flat content string (unchanged, legacy/fallback path).getElements()— An ordered array ofAIAssistantElementobjects representing discrete content blocks (text and cards) in the order the agent produced them. This is the default render source — when present, walk the list left-to-right and render each block in order.
getElements() returns null or is empty (older messages), fall back to getText().
AIAssistantElement
Each element exposes two accessors:| Method | Return Type | Description |
|---|---|---|
getType() | String? | The element’s type: "text", "card", "graph", etc. |
getData() | dynamic | The element’s raw body data. Shape depends on type — String for text, Map<String, dynamic> (with keys card and cardId) for card, raw JSON value for others. |
- Dart
Card Messages (Developer Cards)
Developer card messages are rich, interactive messages (buttons, images, styled layouts) described as JSON and sent via the Platform API or Bubble Builder. The SDK only receives card messages — it does not send them. ACardMessage arrives with category: "card" and is delivered on the onCardMessageReceived callback of the MessageListener.
CardMessage Class
| Method | Return Type | Description |
|---|---|---|
getCard() | Map<String, dynamic>? | The raw card schema payload. Pass directly to the card renderer. |
getText() | String? | Preview text for push notifications and conversation list. |
getFallbackText() | String? | Fallback text from inside the card (card.fallbackText). Used for accessibility or when the renderer fails. |
getTags() | List<String>? | Tags associated with this message. |
- Dart
Card messages are receive-only. They are created and sent exclusively via the Platform (REST) API and Dashboard Bubble Builder. The SDK exposes the card payload raw via
getCard() — an external renderer library (e.g., CometChatCardView) is responsible for drawing the card UI.