LangChain Deep Agents: Batteries-Included Framework for Complex LLM Tasks

Introduction

The previous article "Manus's Core Technology: Planning-with-Files?" drew primarily on insights from Lance Martin, a LangChain engineer. LangChain's open-source Deep Agents framework distills the core patterns behind Manus.

What Is Deep Agents?

Deep Agents is defined as a "batteries-included" agent framework. Unlike traditional approaches that require developers to manually craft prompts, chain tools, and manage context, Deep Agents provides a preset logical architecture.

Design inspiration: Inspired by complex agent applications such as Claude Code, Deep Research, and Manus.

Core philosophy: Enables agents to handle complex cross-file coding, research, and execution tasks through structured planning and context isolation.

Technical foundation: Built on LangGraph (graph execution and state management), LangChain (tool and model integration), and LangSmith (observability and evaluation).

Characteristics of Deep Agents

All agents share the same core algorithm: an LLM that loops calling tools. What makes a "deep" agent different? According to LangChain's Harrison, the distinction lies in four key traits:

Comprehensive system prompts: Deep agents typically have very long system prompts containing detailed instructions on tool usage and few-shot examples for specific situations.

Planning tool: Equipped with a planning tool (e.g., Todo) to decompose complex tasks into steps and track progress. Even if the tool performs no actual operation (like Claude Code's no-op planning tool), it keeps the agent on track over longer horizons.

Sub-agents: Ability to spawn specialized sub-agents for task decomposition. This allows deep diving into specific sub-tasks while isolating context to keep the main agent's context clean.

File system access: Agents can read, write, and modify files. The file system acts as external memory and a shared workspace for collaboration, helping manage the large context accumulated during long runs.

These traits let deep agents go beyond simple "loop-calling-tools" patterns, gaining the ability to handle complex, multi-step tasks and execute goals over extended periods.

Deep Agents Capabilities

Deep Agents provides six core capabilities.

3.1 Planning — Task Planning and Tracking

Provides write_todos and read_todos tools. Implementation resides in langchain.agents.middleware.TodoListMiddleware. This capability lets agents break down complex tasks into steps and continuously track progress.

3.2 Filesystem — File System Operations

Provides ls, read_file, write_file, edit_file, glob, and grep tools. Implementation is in libs/deepagents/deepagents/middleware/filesystem.py. The file system serves as the agent's "external memory" and "shared workspace." Its primary mechanism offloads large contexts (long documents or codebases) to memory or disk storage, preventing LLM context window overflow.

3.3 Shell Access — Shell Command Execution

Provides an execute tool. Implementation is in libs/deepagents/deepagents/backends/local_shell.py. Supports running shell commands on the machine or in a sandboxed environment. The core safety mechanism is sandboxing. DeepAgents follows a "trust the LLM" principle; boundary enforcement happens at the tool or sandbox layer, not by relying on model self-discipline.

3.4 Sub-agents — Sub-agent Delegation

Provides a task tool. Implementation is in libs/deepagents/deepagents/middleware/subagents.py. Agents can spawn specialized sub-agents to handle specific sub-tasks. This achieves context isolation: the main agent delegates messy details to a sub-agent, which "digs deep" in an independent context window, keeping the main context clean.

3.5 Smart Defaults — Intelligent Default Prompts

Provides carefully designed system prompts containing detailed tool-usage instructions and behavioral examples. These prompts teach the model how to behave in specific scenarios and are key to giving agents "deep" execution ability.

Base prompt located in graph.py.

CLI default prompt located in libs/cli/deepagents_cli/default_agent_prompt.md.

3.6 Context Management

Core functions:

Automatic summarization of long conversations.

Saving large tool results to the file system.

Intelligent truncation of tool parameters.

Implementation is in libs/deepagents/deepagents/middleware/summarization.py.

Summary

Deep Agents is a mature architectural template that shifts agent development from "prompt stacking" to "system-level engineering." The author plans to explore implementation details in future articles.

References: 1. https://github.com/langchain-ai/deepagents 2. https://blog.langchain.com/deep-agents/