Subagents: Advanced

This page covers the details of subagents that you need only for special cases. Set up the basics first with Subagents.

Router picks#

The router can return:

• 'main': the parent model answers the turn.
• One agent name.
• An array of names.
• { names, order }: a list with its own order for this turn.
• { steps }: groups that run one after another.

Anywhere a name goes, { name, input } can go too, for example { name: 'pricer', input: { vendor: 'Acme' } }. These rules apply:

• An agent with inputSchema needs an input.
• No other agent takes an input.
• If the pick breaks one of these rules, chat() throws.

One step holds one input for each agent. To give one agent two inputs, put each pick in its own step of { steps }.

Order#

subagents.order is the default order for an array:

• parallel (default): the names start together.
• sequence: the names run one after another. Each later child reads the text of the earlier child.

{ names, order } overrides that default for one turn. If some turns are parallel and some are serial, use it.

{ steps } runs one group, then the next group. Each group has names and an optional order. The next group reads the text from the earlier group. If two agents start together and a later agent must read both, use { steps }:

const plan = {
  steps: [
    { names: ['researcher', 'seo'], order: 'parallel' },
    { names: ['writer'] },
  ],
}

A later user message can also read that child text. The assistant message keeps the name and text of each child, so the next turn still has the notes.

Build the plan with subagentRoute#

subagentRoute makes one yes/no question for each agent and asks Jev for the order. Jev picks parallel when no agent must read text from another agent. The same topic is not a reason to wait. Jev picks sequence only when a later agent must read the earlier text, such as research notes and then a draft.

import { chat, decide, defineAgent, subagentRoute } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'
import { typesafeDecider } from '@tanstack/ai-typesafe'

const researcher = defineAgent({
  name: 'researcher',
  description: 'Looks up facts',
  run: async function* () {},
})
const writer = defineAgent({
  name: 'writer',
  description: 'Writes the post',
  run: async function* () {},
})
const messages = [
  { role: 'user' as const, content: 'Research squids and write an article' },
]

const stream = chat({
  adapter: openaiText('gpt-6.1-sol'),
  messages,
  threadId: 'thread-1',
  subagents: {
    agents: [researcher, writer],
    router: async ({ messages: turnMessages, agents }) => {
      const state = turnMessages.at(-1)
      if (state === undefined) {
        throw new Error('No message')
      }
      const route = subagentRoute(agents)
      const result = await decide({
        adapter: typesafeDecider('jev-latest'),
        state,
        questions: route.questions,
      })
      return route.pick(result)
    },
  },
})

• Pass the agents argument of the router to subagentRoute. Each yes/no question uses the description of that agent.
• If you need different question text, pass when. when must include every agent name.
• pick returns main, one name, { names, order }, or { steps }. The names follow the order of that agents array.
• subagentRoute(agents, { then: ['writer'] }) puts the writer in a later step when it is selected with other agents. The other names start together. The writer runs after them and reads their text. then does not have to list every agent.

Give a routed agent its input#

An agent with inputSchema gets its input from the router:

1. route.needsInput(result) lists the picked agents that need an input, each with its schema.
2. Make each input.
3. Pass the inputs to route.pick(result, { inputs }).

Here the model writes each input, with the schema as outputSchema:

import { chat, decide, defineAgent, subagentRoute } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'
import { typesafeDecider } from '@tanstack/ai-typesafe'
import { z } from 'zod'

const writer = defineAgent({
  name: 'writer',
  description: 'Writes the post',
  run: async function* () {},
})
const pricer = defineAgent({
  name: 'pricer',
  description: 'Compares the plans and prices of one vendor',
  inputSchema: z.object({ vendor: z.string() }),
  run: (ctx) =>
    chat({
      adapter: openaiText('gpt-6.1-sol'),
      messages: [{ role: 'user', content: `Compare the plans of ${ctx.input.vendor}` }],
      threadId: ctx.threadId,
      runId: ctx.runId,
      parentRunId: ctx.parentRunId,
      subagentRunId: ctx.subagentRunId,
      resume: ctx.resume,
    }),
})
const messages = [
  { role: 'user' as const, content: 'What does Acme cost? Write a post about it.' },
]
const threadId = 'thread-1'

const stream = chat({
  adapter: openaiText('gpt-6.1-sol'),
  messages,
  threadId,
  subagents: {
    agents: [pricer, writer],
    router: async ({ messages: turnMessages, agents }) => {
      const route = subagentRoute(agents, { then: ['writer'] })
      const result = await decide({
        adapter: typesafeDecider('jev-latest'),
        state: turnMessages,
        questions: route.questions,
      })
      const inputs = Object.fromEntries(
        await Promise.all(
          route.needsInput(result).map(async ({ name, inputSchema }) => [
            name,
            await chat({
              adapter: openaiText('gpt-6.1-sol'),
              messages: turnMessages,
              threadId,
              outputSchema: inputSchema,
            }),
          ]),
        ),
      )
      return route.pick(result, { inputs })
    },
  },
})

• pick types inputs when you write it as an object, such as { pricer: { vendor: 'Acme' } }. Each key is the name of an agent with inputSchema, and the value has the type of that schema. Object.fromEntries returns any, so the sample above has no such type check.
• If a picked agent with a schema has no input, pick throws Agent "pricer" needs input. Pass it in inputs.
• chat() checks each input against the schema before any child starts. If one input does not match, chat() throws, and no child starts.
• The inputSchema must be a Standard Schema, such as a zod schema. If it is a plain JSON Schema, chat() throws.
• run reads the input as ctx.input. The input stays on the server. It does not go to the client.

The tool input#

Without a router, every agent tool takes the fields of the agent, plus two more:

• The fields of inputSchema, for an agent with one. The tool checks them against the schema.
• prompt: the task as text, for an agent without inputSchema. The child gets it as the last user message.
• sessionId: the subagentRunId of an earlier result. The call continues that child. The tool has this field only when withPersistence is on the parent chat().
• background: true starts the child and returns its subagentRunId at once. The tool has this field only when a host that runs agents for you, such as a harness, turns it on. maxConcurrent counts that child until it ends.

When the run uses sessionId or background, that name is reserved. If an inputSchema has a field with that name, chat() throws before the run starts.

Wrong calls#

A wrong call goes back to the model as a tool error, and the run continues. For example:

• Input validation failed for tool researcher: ...: the fields do not match inputSchema.
• Unknown sessionId "...": no child of this thread has that id. A thread can continue only its own children.
• Session "..." belongs to agent "writer", not "researcher".: the call must use the tool of the agent that ran the child.
• background needs a host that runs children in the background.

Return a value, or call any activity#

A child does not have to be a chat. It can make an image, or return a plain value that the parent model reads as the tool result. Call the activity on ctx and return what it gives you:

import { chat, defineAgent } from '@tanstack/ai'
import { openaiImage, openaiText } from '@tanstack/ai-openai'
import { z } from 'zod'

const heroImage = defineAgent({
  name: 'heroImage',
  description: 'Makes a hero image for a post',
  produces: 'image',
  inputSchema: z.object({ prompt: z.string() }),
  run: (ctx) =>
    ctx.generateImage({
      adapter: openaiImage('gpt-image-2.5-sunburst'),
      prompt: `${ctx.input.prompt}. Brand colors: blue and white.`,
      size: '1536x1024',
    }),
})

const stream = chat({
  adapter: openaiText('gpt-6.1-sol'),
  messages: [{ role: 'user', content: 'Make a hero image about squids' }],
  threadId: 'thread-1',
  subagents: { agents: [heroImage] },
})

What the activities on ctx do:

• ctx.chat, ctx.generateImage, ctx.generateVideo, ctx.generateSpeech, and the other activities on ctx take the same options as the plain functions. They fill in the thread id, a run id, and the abort signal.
• ctx.chat also uses the parent conversation (ctx.messages) when you do not pass messages.
• Each ctx.chat call after the first one gets its own run id.
• A child that ctx.chat starts gets the prompt cache retention of the parent call, unless the call sets its own promptCache.

What happens to the result:

• run can return a promise of any value. The value arrives on SUBAGENT_FINISHED.result, and the parent model gets it as the tool result.
• A string result also streams as the text of the child, so the card shows it. The parent model gets the full string.
• The result of ctx.chat({ stream: false }) streams like the chat stream. Its text, usage, and interrupts reach the parent.
• A very long string in the result (for example a base64 image) reaches the parent model as a short note, [omitted 5000 characters]. The full value stays on SUBAGENT_FINISHED.result.

produces says what the agent makes ('image', 'text', and more). It does not change how the agent runs.

When you call the plain chat() instead, spread ctx.forward. It passes the ids and the abort controller of the child in one line:

const researcher = defineAgent({
  name: 'researcher',
  description: 'Looks up facts',
  run: (ctx) =>
    chat({
      adapter: openaiText('gpt-6.1-sol'),
      messages: ctx.messages,
      ...ctx.forward,
    }),
})

Limit details#

The whole tree shares one budget. A child that calls ctx.chat({ subagents }) passes the budget on, so a child cannot reset it.

• Set limits on the first chat(). A nested chat in a child uses that budget, so its own limits do not apply.
• A child that a router picks does not count against the limits. The children it starts do count, one level deeper.

A refused start reaches the model as a tool error:

Keys for agents#

An agent that calls a model needs a provider key. An adapter such as openaiText reads its env var, for example OPENAI_API_KEY.

To use the key of the user, pass it on subagents.keys. The agent then builds its adapter with ctx.keys. See Keys inside agents.

Middleware#

A child is its own chat() call. Put the middleware of the child in that call. The middleware list of the parent does not reach the child.

The middleware context of a child has these fields:

• subagentRunId: tells a child run from a top-level run. Use it to link a child trace to its card.
• subagentName: the name of the agent, when the child calls ctx.chat.
• parentSubagentRunId: for a nested child, the id of the child that started it.

When a resume answers an interrupt from an agent tool call, the middleware context of the parent has childResumeIds. These are the interrupt ids in resume that the child answers. The parent run does not answer them. When no child resumes, childResumeIds is undefined.

On a routed turn where a child runs and main does not, the middleware of the parent does not run. Only withPersistence records that turn. A turn where main runs, including a handoff, runs the middleware of the parent as usual.

Run a child from code#

A server job can run an agent with no parent chat turn. If the job must read a pause or the token counts, pass a sink.

import {
  chat,
  createSubagentId,
  createSubagentSink,
  defineAgent,
  runAgentStream,
} from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'

const researcher = defineAgent({
  name: 'researcher',
  description: 'Looks up facts and sources',
  run: (ctx) =>
    chat({
      adapter: openaiText('gpt-6.1-sol'),
      messages: ctx.messages,
      ...ctx.forward,
    }),
})

const subagentRunId = createSubagentId()
const sink = createSubagentSink()
for await (const chunk of runAgentStream(
  researcher,
  {
    input: undefined,
    messages: [{ role: 'user', content: 'Find three facts about squids' }],
    threadId: 'thread-1',
    runId: `job-1:${subagentRunId}`,
    parentRunId: 'job-1',
    subagentRunId,
  },
  sink,
)) {
  if (chunk.type === 'TEXT_MESSAGE_CONTENT') console.log(chunk.delta)
}

The sink holds the pause and the token counts:

• sink.interrupts has the interrupt objects. The finished event has interruptIds only.
• sink.usage has one entry for each model call.

After the loop, read sink.interrupts to show the pause.

Some things that chat({ subagents }) does for you, runAgentStream does not do:

• It does not check input. If the agent has an inputSchema, check the input before the call.
• A failure ends as SUBAGENT_ERROR. The stream omits RUN_STARTED, MESSAGES_SNAPSHOT, and RUN_FINISHED.

Sandbox#

withSandbox keys the workspace by threadId. Set subagents.sandbox so the child gets the right thread id. Pass ctx.threadId into the child chat().

• 'own' (default): the child thread id is ${parentThreadId}:${name}. The child gets its own workspace. Parallel children each get a workspace.
• 'inherit': the child thread id is the parent thread id. The child uses the parent workspace again when lifecycle.reuse is 'thread'.

If sandbox is 'inherit' and the router returns two or more names, chat() throws before the start. Otherwise two children write the same files at the same time.

import { chat, defineAgent } from '@tanstack/ai'
import { grokBuildText } from '@tanstack/ai-grok-build'
import {
  defineSandbox,
  defineWorkspace,
  withSandbox,
} from '@tanstack/ai-sandbox'
import { dockerSandbox } from '@tanstack/ai-sandbox-docker'

const messages = [{ role: 'user' as const, content: 'Fix the tests' }]
const threadId = 'parent-thread'

const repoSandbox = defineSandbox({
  id: 'repo-agent',
  provider: dockerSandbox({ image: 'node:22' }),
  workspace: defineWorkspace({
    source: { type: 'none' },
    packageManager: 'pnpm',
  }),
  lifecycle: { reuse: 'thread' },
})

const coder = defineAgent({
  name: 'coder',
  description: 'Edits the repo in a sandbox',
  run: (ctx) =>
    chat({
      adapter: grokBuildText('grok-build'),
      messages: ctx.messages,
      threadId: ctx.threadId,
      runId: ctx.runId,
      parentRunId: ctx.parentRunId,
      subagentRunId: ctx.subagentRunId,
      resume: ctx.resume,
      middleware: [withSandbox(repoSandbox)],
    }),
})

const ownStream = chat({
  adapter: grokBuildText('grok-build'),
  messages,
  threadId,
  middleware: [withSandbox(repoSandbox)],
  subagents: {
    agents: [coder],
    sandbox: 'own',
    strategy: 'exclusive',
    router: () => 'coder',
  },
})

const inheritStream = chat({
  adapter: grokBuildText('grok-build'),
  messages,
  threadId,
  middleware: [withSandbox(repoSandbox)],
  subagents: {
    agents: [coder],
    sandbox: 'inherit',
    strategy: 'exclusive',
    router: () => 'coder',
  },
})

• ownStream runs the child with threadId parent-thread:coder.
• inheritStream runs the child with threadId parent-thread, so reuse: 'thread' shares the parent workspace.

See Sandboxes for withSandbox and lifecycle.reuse.

Client details#

The nested type: 'subagent' part and useChat().subagents[i] are the same live object. When you call stop() on either one, this happens:

• The client sets that child to error.
• The client aborts the current parent run.
• The client ignores later events for that id, and for its nested children.

part.subagent.messages holds everything the child did:

• text and reasoning
• tool calls and tool results
• the approval state on the tool calls of the child
• nested children, as their own subagent parts

part.subagent.status is 'running', 'finished', 'error', or 'suspended'.

Type the child parts#

Pass the agents that you give to chat() to useChat too. The hook uses them for types only. It does not call run.

import { defineAgent } from '@tanstack/ai'
import { fetchServerSentEvents, useChat } from '@tanstack/ai-react'

const researcher = defineAgent({
  name: 'researcher',
  description: 'Looks up facts',
  run: async function* () {},
})
const writer = defineAgent({
  name: 'writer',
  description: 'Drafts posts',
  run: async function* () {},
})

function Desk() {
  const chat = useChat({
    connection: fetchServerSentEvents('/api/chat'),
    threadId: 'support-chat',
    subagents: [researcher, writer],
  })
  const part = chat.messages[0]?.parts[0]
  if (part?.type === 'subagent' && part.subagent.name === 'researcher') {
    return part.subagent.status
  }
  return null
}

part.subagent.name is 'researcher' | 'writer'. After you check the name, the messages of that child use the tools and the output schema of that agent.

When run imports server code, do not import the agent into the browser. Do these steps:

1. Declare the agent once in a shared file, with its name, description, and tool definitions from toolDefinition.
2. On the server, pass defineAgent({ ...researcher, run }) to chat().
3. On the client, pass the declaration.

Draw cards with createChatHook#

If you want the factory to draw the cards, use createChatHook from @tanstack/ai-react/ui:

• Pass the agents that you give to chat() as options.subagents. The client reads the name, tools, interrupts, and outputSchema of each agent for types. It does not call run.
• Register subagentsComponents for each agent name. Those components receive SubagentProps and Parts.
• Render <Messages />. The subagent card is a part of the assistant message. <Subagents /> draws that same card for the live list. Pick one place for the card.

import { defineAgent } from '@tanstack/ai'
import { fetchServerSentEvents } from '@tanstack/ai-react'
import { createChatHook } from '@tanstack/ai-react/ui'
import type { LayoutProps, SubagentProps } from '@tanstack/ai-react/ui'

const researcher = defineAgent({
  name: 'researcher',
  description: 'Looks up facts',
  run: async function* () {},
})
const writer = defineAgent({
  name: 'writer',
  description: 'Drafts posts',
  run: async function* () {},
})

const chatOptions = {
  connection: fetchServerSentEvents('/api/chat'),
  subagents: [researcher, writer],
}

function SubagentCard({
  subagent,
  Parts,
}: SubagentProps<typeof chatOptions, 'researcher' | 'writer'>) {
  return (
    <section>
      <strong>{subagent.name}</strong>
      <span>{subagent.status}</span>
      {subagent.status === 'running' ? (
        <button type="button" onClick={() => subagent.stop?.()}>
          Stop
        </button>
      ) : null}
      <Parts />
    </section>
  )
}

const { useAppChat, useChatContext } = createChatHook({
  options: chatOptions,
  components: {
    layout: ({ Messages, Input }: LayoutProps<typeof chatOptions>) => (
      <main>
        <Messages />
        <Input />
      </main>
    ),
    message: ({ Parts }) => <article><Parts /></article>,
    input: function Input() {
      const chat = useChatContext()
      return (
        <form
          onSubmit={(event) => {
            event.preventDefault()
            const field = event.currentTarget.elements.namedItem('message')
            if (!(field instanceof HTMLInputElement)) return
            void chat.sendMessage(field.value)
            field.value = ''
          }}
        >
          <input name="message" />
          <button type="submit">Send</button>
        </form>
      )
    },
  },
  partsComponents: {
    text: ({ part }) => <p>{part.content}</p>,
    fallback: () => null,
  },
  subagentsComponents: {
    researcher: SubagentCard,
    writer: SubagentCard,
  },
})

export function ChatScreen() {
  const chat = useAppChat({ threadId: 'support-chat' })
  return <chat.AppChat />
}

By default, the child messages use the same parts components as the parent. To change that for one card, see Style one child's parts.

Show the brief on a card#

When an agent has inputSchema, the card keeps the checked input that the parent model wrote, as part.subagent.input. Put inputSchema in the shared declaration, and pass the declaration to useChat to type it:

import { z } from 'zod'

export const researcher = {
  name: 'researcher',
  description: 'Researches one focused question and returns sourced findings',
  inputSchema: z.object({ task: z.string() }),
} as const

import { fetchServerSentEvents, useChat } from '@tanstack/ai-react'
import { researcher } from './agents'

export function Briefs() {
  const { messages } = useChat({
    connection: fetchServerSentEvents('/api/chat'),
    threadId: 'support-chat',
    subagents: [researcher],
  })
  return messages.flatMap((message) =>
    message.parts.map((part) =>
      part.type === 'subagent' ? (
        <p key={part.subagent.id}>{part.subagent.input?.task}</p>
      ) : null,
    ),
  )
}

• part.subagent.input is the same value as ctx.input in run, with the output type of the schema. Here it is { task: string } | undefined.
• It is undefined for an agent without inputSchema.
• A router input stays on the server, so a child that a router started has no part.subagent.input.

Style one child's parts#

The reasoning and tool calls of the researcher use the root widgets by default. To make them look different on the researcher card only, pass widgets to the Parts of that card:

import { toolDefinition } from '@tanstack/ai'
import { fetchServerSentEvents } from '@tanstack/ai-react'
import { createChatHook, ThinkingPart } from '@tanstack/ai-react/ui'
import type { SubagentPartsProps, SubagentProps } from '@tanstack/ai-react/ui'
import { z } from 'zod'

const lookupWikipedia = toolDefinition({
  name: 'lookupWikipedia',
  description: 'Get the Wikipedia summary of one topic',
  inputSchema: z.object({ title: z.string() }),
  outputSchema: z.object({ extract: z.string() }),
})

// Shared with the server, which passes defineAgent({ ...researcher, run }).
const researcher = {
  name: 'researcher',
  description: 'Looks up facts',
  tools: [lookupWikipedia],
} as const

const chatOptions = {
  connection: fetchServerSentEvents('/api/chat'),
  subagents: [researcher],
}

type ResearcherWidgets = SubagentPartsProps<typeof chatOptions, 'researcher'>

const researcherParts: ResearcherWidgets['partsComponents'] = {
  thinking: ({ part }) => (
    <ThinkingPart content={part.content} className="research-notes" />
  ),
}

const researcherTools: ResearcherWidgets['toolsComponents'] = {
  lookupWikipedia: ({ part }) => (
    <details>
      <summary>
        {part.input?.title} ({part.state})
      </summary>
      {part.output?.extract}
    </details>
  ),
}

function Researcher({
  Parts,
}: SubagentProps<typeof chatOptions, 'researcher'>) {
  return (
    <section>
      <Parts partsComponents={researcherParts} toolsComponents={researcherTools} />
    </section>
  )
}

export const { useAppChat } = createChatHook({
  options: chatOptions,
  components: {
    layout: ({ Messages }) => (
      <main>
        <Messages />
      </main>
    ),
    message: ({ Parts }) => (
      <article>
        <Parts />
      </article>
    ),
  },
  partsComponents: {
    text: ({ part }) => <p>{part.content}</p>,
    thinking: ({ part }) => <ThinkingPart content={part.content} />,
    fallback: () => null,
  },
  subagentsComponents: { researcher: Researcher },
})

• An entry on Parts replaces the root entry with the same key, for this card and for the children nested in it.
• A key that you do not set uses the root entry. Here, text still uses the root widget.
• The keys and props of toolsComponents come from the tools of the agent. part.input and part.output are typed, and a tool name that the agent does not have is a type error.
• A tool call with no entry at any level renders nothing.
• An approval for a child tool arrives as interrupt on that tool widget, typed from the tool. The root interruptsComponents also accepts the approval tools of the child and the ids in the interrupts of the agent.
• An approval for a tool that only a card registers also shows in the root <Interrupts /> list.

Define these maps outside the component. A new object on each render makes every part in the card render again.

Devtools#

In the AI devtools, the Conversation tab shows each child as a card with its name and status:

• The card holds the steps of the child, drawn the same way as the steps of the parent.
• A new step starts after each tool result.
• The User view on the right shows the text and tool outputs of the child.

When the server events reach the devtools (the devtools() plugin from @tanstack/devtools-vite), the steps of the child are its real server iterations, with the model, the system prompts, and the token usage. Without the plugin, the steps come from the messages in the browser.

See Stream Events for SUBAGENT_* and subagentRunId.