Stream an agent response

Use agent.prompt(...).stream() when the UI should update while the agent is still working.

When to use this

Use agent streaming when you need progressive text, tool-call visibility, final usage, or error events from one agent run.

Prerequisites

Send a prompt with send() first so you know the agent works without streaming.

Stream an agent run

for await (const event of agent.prompt("Draft a reply to this customer.").stream()) {
  if (event.type === "text_delta") {
    process.stdout.write(event.delta);
  }

  if (event.type === "final") {
    console.log(event.usage);
  }
}

Event types

Agent streams include runtime events, not only model text:

• turn_start: a model turn is starting.
• text_delta: visible assistant text.
• reasoning_delta: provider reasoning text when the model emits it.
• tool_call: the model requested a tool call.
• tool_result: your app returned a tool result.
• turn_end: a model turn completed with a provider response.
• agent_tool_event: a nested agent emitted an event while running as a tool.
• final: the run id, completed output, usage, messages, and trace.
• error: an error from the active run.

Completion stream vs agent stream

createCompletionStream streams one model call. agent.prompt(...).stream() streams the whole agent run, including tools and final runtime metadata.

Check yourself

Run the loop and confirm you receive text_delta events before the final event.

Next

Add tools so the agent can call product actions.

Add tools