Subagent

Delegates a self-contained task to a child agent that runs in its own context and returns a final report. Use it when a subtask would otherwise flood the parent’s context with intermediate work and you only need the conclusion. The model writes the task each time it calls the tool; the child cannot ask follow-up questions, so the task must include all the context it needs.

The make_subagentmakeSubagent factory uses axis policies to control what the model can configure on each child. The configurable axes are instructions, tools, model, and context. The parent also passes a task to the child.

By default, the child sees only the task (context=Fixed("none")context: new Fixed('none')), but you can share the parent’s conversation by setting context to a Choice over "none", "all" (full history including tool calls), and "no_tools" (text turns only). When a non-"none" option is available, the model can also pass last_messages to limit how many parent messages to share.

Supported in: Python, TypeScript (Node.js, browsers).

Quickstart example:

TypeScript
Python

import { Agent } from '@strands-agents/sdk'
import { subagent } from '@strands-agents/sdk/vended-tools/subagent'

const agent = new Agent({
  systemPrompt: 'You are a manager.',
  tools: [subagent],
})
await agent.invoke('Research the latest TypeScript 5.x features and summarize them.')

from strands import Agent
from strands.vended_tools.subagent import subagent

agent = Agent(
    system_prompt="You are a manager.",
    tools=[subagent],
)
agent("Research the latest Python 3.13 features and summarize them.")

Example — full control with make_subagentmakeSubagent:

TypeScript
Python

import { Agent } from '@strands-agents/sdk'
import { Choice, Option, Preset } from '@strands-agents/sdk/multiagent'
import { makeSubagent } from '@strands-agents/sdk/vended-tools/subagent'

const subagent = makeSubagent({
  presets: {
    researcher: new Preset({
      instructions: 'You research topics thoroughly.',
      description: 'deep research on a topic',
    }),
    reviewer: new Preset({
      instructions: 'You review code for correctness and style.',
      description: 'code review',
    }),
  },
  model: new Choice([
    new Option('fast', 'global.anthropic.claude-sonnet-4-6', 'quick tasks'),
    new Option('deep', 'global.anthropic.claude-opus-4-6-v1', 'hard problems'),
  ]),
  context: new Choice(['none', 'all', 'no_tools']),
  tools: new Choice(['read', 'shell', 'write'], true),
  maxDepth: 3,
})
const agent = new Agent({ tools: [subagent] })
await agent.invoke('Review the changes in src/main.ts for correctness.')

from strands import Agent
from strands.multiagent.spec import Choice, Option, Preset
from strands.vended_tools.subagent import make_subagent

subagent = make_subagent(
    presets={
        "researcher": Preset(
            instructions="You research topics thoroughly.",
            description="deep research on a topic",
        ),
        "reviewer": Preset(
            instructions="You review code for correctness and style.",
            description="code review",
        ),
    },
    model=Choice([
        Option("fast", "global.anthropic.claude-sonnet-4-6", "quick tasks"),
        Option("deep", "global.anthropic.claude-opus-4-6-v1", "hard problems"),
    ]),
    context=Choice(["none", "all", "no_tools"]),
    tools=Choice(["read", "shell", "write"], multiple=True),
    max_depth=3,
)
agent = Agent(tools=[subagent])
agent("Review the changes in src/main.py for correctness.")

Full API reference: TypeScript