Skip to main content

Documentation Index

Fetch the complete documentation index at: https://relevanceai.com/docs/llms.txt

Use this file to discover all available pages before exploring further.

Messages are how your application and an agent exchange information inside a task. This page covers sending messages, the different kinds of incoming messages, and how to handle them.

​
Send a message

Both agents and workforces use sendMessage. It accepts a string and returns a Task representing the conversation thread. 
const task = await agent.sendMessage("What can you help me with?");
The returned task begins processing immediately. The promise resolves as soon as the message is sent — not when the agent responds. To receive the response, listen for events on the task. See Tasks for event handling.

​
Continue a conversation

To send a follow-up message within the same conversation, pass the existing task as the second argument: 
const task = await agent.sendMessage("Hello");

// Later, continue the same conversation
await agent.sendMessage("Tell me more", task);
Each call with an existing task appends to the conversation history. The agent has full context of the prior messages.

​
File attachments

Attach files by passing an array of File objects as the second argument. The SDK handles the upload. 
const pdf = new File([bytes], "contract.pdf", {
  type: "application/pdf",
});

const task = await agent.sendMessage("Summarize this contract", [pdf]);
Multiple files in a single message: 
const task = await agent.sendMessage(
  "Compare these two documents",
  [fileA, fileB]
);
To send attachments while continuing a conversation, pass the attachments array and the task: 
await agent.sendMessage(
  "Here is an updated version",
  [updatedFile],
  task
);
Pre-uploaded attachments (objects with fileName and fileUrl fields) can be mixed in with File objects in the same array.
Workforce tasks don’t support file attachments. Only agent tasks accept files.

​
Handle incoming messages

Messages arrive through the "message" event on a task. Each event’s detail includes a message property. Use the type guard methods to determine the kind and access the right fields. 
task.addEventListener("message", ({ detail }) => {
  const { message } = detail;

  if (message.isAgent()) {
    console.log("Agent:", message.text);
  } else if (message.isTool()) {
    console.log("Tool:", message.status);
  } else if (message.isThinking()) {
    console.log("Thinking:", message.text);
  } else if (message.isTyping()) {
    console.log("Typing:", message.text);
  }
});
The type guards narrow the TypeScript type — properties specific to each message kind become available after the check.

​
Agent responses

When message.isAgent() returns true, the message is an AgentMessage carrying the agent’s response text. 
if (message.isAgent()) {
  console.log(message.text);
  console.log(message.agentId);
}

​
Tool executions

When message.isTool() returns true, the message is a ToolMessage representing a tool or subagent execution inside the agent’s pipeline. 
if (message.isTool()) {
  console.log("Tool:", message.tool?.name);
  console.log("Status:", message.status);
}
The status field tracks where the tool is in its lifecycle:
StatusMeaning
pendingScheduled, not yet started
runningCurrently executing
completedFinished successfully
errorFailed with an error
cancelledExecution was cancelled

​
Read tool output

Once a tool completes, its output is available: 
if (message.isTool() && message.status === "completed") {
  console.log(message.output);
}

​
Detect subagents

A tool message may represent a subagent rather than a standalone tool. Use isSubAgent to check, and subAgentTaskId to access the subagent’s task: 
if (message.isTool() && message.isSubAgent()) {
  console.log("Sub-agent task:", message.subAgentTaskId);
}

​
Tool errors

Check for errors on a tool message with hasErrors: 
if (message.isTool() && message.hasErrors()) {
  for (const error of message.errors) {
    console.error(error.stepName, error.message);
  }
}

​
User messages

User messages represent input sent by the application or end user. They show up in the message history alongside agent responses. 
task.addEventListener("message", ({ detail }) => {
  const { message } = detail;

  if (message.isUser()) {
    console.log("User:", message.text);

    if (message.hasAttachments()) {
      for (const attachment of message.attachments) {
        console.log("File:", attachment.fileName);
      }
    }
  }
});
isTrigger returns true for the first message in a conversation — the one that created the task: 
if (message.isUser() && message.isTrigger()) {
  console.log("Conversation started with:", message.text);
}

​
Error handling

Agent errors are delivered through the "error" event on the task. The event detail contains an AgentErrorMessage with one or more error strings. 
task.addEventListener("error", ({ detail }) => {
  const { message } = detail;
  console.error("Last error:", message.lastError);

  // All errors in this cycle
  for (const error of message.errors) {
    console.error(error);
  }
});
Error events mean the agent ran into a problem during execution. The task may transition to an "error" or "action" status depending on the nature of the failure. See Tasks for status details.

​
Workforce messages

Workforce tasks include two additional message types for multi-agent coordination:
  • WorkforceAgentMessage — an agent within the workforce is executing a subtask. Includes details about which agent is running and the state of its work.
  • WorkforceAgentHandoverMessage — work is being delegated from one agent to another within the workforce. Includes the trigger message and details about the receiving agent.
These messages appear alongside standard agent and tool messages in the event stream. See Workforces for workforce-specific behavior.

​
Streaming messages

Two message types represent real-time incremental output from the agent:
  • ThinkingMessage — the agent’s intermediate reasoning as it processes.
  • TypingMessage — the agent’s response text as it’s being generated.
Both are identified with the isThinking and isTyping type guards shown above. For streaming behavior and live typing indicators, see Streaming.