Deep Agents

Deep Agents

When to Use This Skill

Core Expertise

Deep Agents (deepagents) is a TypeScript library for building sophisticated AI agents:

• Built on LangGraph with planning and decomposition
• File system context management (prevents token overflow)
• Subagent delegation for focused exploration
• Persistent memory across conversations
• Modeled after Claude Code and Deep Research patterns

The package name on npm is deepagents (one word, unscoped). The source lives at langchain-ai/deepagentsjs.

Installation

# Install Deep Agents
npm install deepagents

# Add a model provider (pick the one matching your model)
npm install @langchain/openai  # or @langchain/anthropic, @langchain/google-genai

deepagents declares langsmith as a peer dependency (for tracing) and builds on @langchain/langgraph + @langchain/core, which are pulled in transitively.

Basic Agent Setup

createDeepAgent() returns a compiled LangGraph graph. The model can be a provider-prefixed string (e.g. "openai:gpt-5") or a model instance.

import { createDeepAgent } from "deepagents";
import { ChatOpenAI } from "@langchain/openai";

const model = new ChatOpenAI({
  model: "gpt-5",
  temperature: 0,
});

const agent = createDeepAgent({
  model,
  systemPrompt: `You are a research assistant.
    Break complex questions into steps using write_todos.
    Use read_file and write_file to manage context.`,
});

const result = await agent.invoke({
  messages: [{ role: "user", content: "Research X and summarize" }],
});

For browser or Node-explicit builds, import the backend-scoped entrypoints:

import { createDeepAgent, StateBackend } from "deepagents/browser";
import { createDeepAgent, FilesystemBackend } from "deepagents/node";

Built-in Tools

Deep Agents ships these tools automatically: write_todos, ls, read_file, write_file, edit_file, glob, grep, and task.

Planning Tools

// write_todos - Task decomposition (available automatically)

// The agent uses it to plan:
// write_todos([
//   { task: "Search for X", status: "pending" },
//   { task: "Analyze results", status: "pending" },
//   { task: "Write summary", status: "pending" },
// ])

File System Tools

// Built-in tools for context management

// ls        - List directory contents
// read_file - Read file content
// write_file - Write/create files
// edit_file - Modify existing files
// glob      - Match files by pattern
// grep      - Search file contents

// The agent stores intermediate results in files
// to prevent context overflow.

Subagent Delegation

// task - Spawn a focused subagent with an isolated context window

// The parent agent delegates:
// task({
//   description: "Research pricing models",
//   subagent_type: "research-agent",
// })

// The subagent runs independently and returns results.

Agentic Optimizations

Quick Reference

Agent Methods

Built-in Tools

Config Keys

Environment Variables

For custom tools, persistence, full configuration options, multi-agent subagent patterns, context-management strategy, streaming, and the Claude Code comparison, see REFERENCE.md.