Agents

Streaming answers

An agent writes its answer once; every person, device and other agent reads it live.

An agent writes its answer once. The person who asked reads it on their phone, a teammate on their laptop, a supervisor agent in the background, all live, all from the same stream. If one of them reconnects, it resumes from its last position; if one arrives late, it reads from the start.

Write#

const answer = pulse.stream({ streamId: `answer:${taskId}`, preset: 'llm_agent_run' })try {  for await (const text of model.stream(prompt)) await answer.write({ kind: 'text_delta', text })  await answer.end()} catch (err) {  await answer.write({ kind: 'text_delta', text: '\n\n(The answer stopped early.)' })  await answer.end()}

Hand work to another agent#

Because answers are streams, a second agent can read the first one’s output while it is being written and start on its part early: a reviewer that checks a draft as it forms, or a translator that follows a live answer.

answerStream.subscribe((event) => reviewer.feed(event))

Long runs#

Time is free: an agent may stream for an hour, from a queue worker or a serverless job that ends and resumes. The answer counts as one message (again past 10,000 events or 10 MiB). See AI streaming for the details and the AI SDK adapter.