Is it possible to stream text back from a tool call?

[matthew-gill]

I'm using streamText to power a chat app. For one of my tools, I want to hand off to another LLM to generate the tool response.
I'd like to be able to stream the output of the tool to the front end as it's generating, but I can't seem to get it to work.

I've created an example below just to illustrate the issue (I know in this case, a tool call is unnecessary).
Obviously I can await the response within the tool call, but there's a delay while the LLM is generating.
I'd like to avoid a client side request if at all possible.

export async function POST(req: Request) {
  const { messages } = await req.json();

  const result = streamText({
    model: vertexAnthropic('claude-3-7-sonnet@20250219'),
    system: `The user will give you a topic, call the generatePoem tool with the topic`,
    messages,
    maxSteps: 15,
    tools: {
        generatePoem: {
            description: 'Generate poetry',
            parameters: z.object({topic: z.string()}),
            execute: async ({ topic }) => {
                // This part doesn't work
                return await streamText({
                    model: vertexAnthropic('claude-3-7-sonnet@20250219'),
                    prompt: `Write a poem about ${topic} in the style of a pirate`
                })
            }
        }
    }
  });

  return result.toDataStreamResponse();
}

Any ideas if it's possible to stream the result back to the user? TIA!

[matthew-gill]

Update - so what I've ended up doing is creating a set of "global tools" that all my agents have access to.
One tool is called switch agent.

It's a tool call which has a description made up of the different streamText "agents".
It sets a metadata value on the chat called agent and returns json like:

{
     agent: "different_agent",
     nextStep: "fulfil the user's objective"
}

Then, when loading my chat, I see if there is an agent value set against it and dynamically switch out the system prompt, tools and messages with that of my agent.

With maxSteps set to quite high, this seems to work well! You might need some prompt tweaks, e.g. "Return to agent abc upon completion of your task"

[js0926861]

@matthew-gill I need something similar as I try to build Agent Swarm. Are you able to share sample code?

[Fen747]

I'm interested to hear more too :)

[patelvivekdev]

In V4, you can use createDataStreamResponse and update the dataStream in tool calls.

You can also use data parts (V5)
https://github.com/nicoalbanese/ai-sdk-alpha-hacking-data-parts/blob/after/app/api/chat/route.ts

[raythurnvoid]

This is cool thanks for sharing!