Graph

Lets your agent design and run a team of sub-agents as a Graph. The model supplies a task, a list of nodes (each with a unique name and its own instructions), and edges that order them. Nodes with no upstream nodes start in parallel, and each node receives its upstream outputs.

Nodes inherit the caller’s model and tools, including MCP tools, except graph tools. Child interrupts pause the tool call and resume with the user’s response. An interrupted graph is saved in the parent agent’s interrupt state, so it can resume after the parent is restored from a snapshot or session. By default, a call times out after 15 minutes and each node after 5 minutes.

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

Example:

TypeScript
Python

import { Agent } from '@strands-agents/sdk'
import { graph } from '@strands-agents/sdk/vended-tools/graph'
import { httpRequest } from '@strands-agents/sdk/vended-tools/http-request'

const agent = new Agent({ tools: [graph, httpRequest] })
await agent.invoke(
  'Research three competing vector databases in parallel, then write a comparison table.'
)

Use makeGraph to set limits (maxNodes, maxEdges, maxSteps, and timeout and nodeTimeout in milliseconds), build each node agent yourself with builder, or control what the model configures on each node with axis policies over instructions, tools, and model.

import { Agent } from '@strands-agents/sdk'
import { Choice, Preset } from '@strands-agents/sdk/multiagent'
import { makeGraph } from '@strands-agents/sdk/vended-tools/graph'
import { fileEditor } from '@strands-agents/sdk/vended-tools/file-editor'
import { httpRequest } from '@strands-agents/sdk/vended-tools/http-request'

const team = makeGraph({
  presets: {
    researcher: new Preset({
      instructions: 'Gather facts from the web and cite every source.',
      tools: ['http_request'],
      description: 'Finds and summarizes sources.',
    }),
    writer: new Preset({
      instructions: 'Turn upstream findings into clear prose.',
      tools: ['file_editor'],
      description: 'Drafts the final document.',
    }),
  },
  tools: new Choice(['http_request', 'file_editor'], true),
  maxNodes: 6,
  maxEdges: 10,
  timeout: 600_000,
  nodeTimeout: 120_000,
})
const agent = new Agent({ tools: [team, httpRequest, fileEditor] })

from strands import Agent
from strands.vended_tools import graph, http_request

agent = Agent(tools=[graph, http_request])
agent("Research three competing vector databases in parallel, then write a comparison table.")

Use make_graph to set limits (max_nodes, max_edges, max_node_executions, and execution_timeout and node_timeout in seconds), build each node agent yourself with builder, or control what the model configures on each node with axis policies over instructions, tools, and model. A running graph cannot be cancelled from the parent agent.

from strands import Agent
from strands.multiagent import Choice, Preset
from strands.vended_tools import file_editor, http_request, make_graph

team = make_graph(
    presets={
        "researcher": Preset(
            instructions="Gather facts from the web and cite every source.",
            tools=["http_request"],
            description="Finds and summarizes sources.",
        ),
        "writer": Preset(
            instructions="Turn upstream findings into clear prose.",
            tools=["file_editor"],
            description="Drafts the final document.",
        ),
    },
    tools=Choice(["http_request", "file_editor"], multiple=True),
    max_nodes=6,
    max_edges=10,
    execution_timeout=600,
    node_timeout=120,
)
agent = Agent(tools=[team, http_request, file_editor])

Full API reference: TypeScript