hwchase17

deepagents: Build Agents with Planning and Tooling Built In

Deep Agents is a Python agent harness for building LLM agents with planning, delegation, file access, and context management built in. It suits developers who need a more capable starting point than a minimal agent loop, while retaining control over models and tools.

View on GitHub Compare
Stars
29.9k
Forks
4.2k
Stars, last 30 days
+985

Our take

Health score

84/100

How it is scored
Activity
90
Community
79
Issues
83
Pull requests
83

+985 stars in the last 30 days

  • Merges more pull requests than 95% of the projects we track
  • Releases more often than 93% of the projects we track
  • Closes issues faster than 81% of the projects we track

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

31,379

Forks

4,429

Change since Nov 17, 2026: +37 stars·+2 forks

Overview

Deep Agents is an opinionated harness for creating LLM agents that handle multi-step work. It bundles capabilities such as planning, file operations, sub-agent delegation, and context management so developers do not need to assemble each one from scratch.

Use it when a ready-to-extend agent structure fits better than a minimal agent loop. It is built on LangGraph and supports tool-calling models; developers can customize its components or use LangChain and LangGraph directly for lighter or more custom orchestration.

Key Features

  • Delegate work to sub-agents with isolated context windows.
  • Read, write, edit, and search files using pluggable backends.
  • Manage long conversations by summarizing threads and offloading tool outputs.
  • Run shell commands through a chosen sandbox.
  • Persist memory across sessions with configurable state and store backends.
  • Add human approval, editing, or rejection steps for tool calls.
  • Load reusable skills on demand and connect custom tools or MCP servers.

Use Cases

  • Research assistants that plan multi-step investigations and produce file-based summaries.
  • Coding or operations agents that need file and shell tools, with execution boundaries enforced by a sandbox.
  • Internal assistants that delegate complex tasks to specialized sub-agents.
  • Long-running workflows where conversation context or useful state must persist across sessions.

Getting Started

Install with uv add deepagents. See the README and documentation for setup and configuration details.

Alternatives

  • mcp-agent: mcp-agent centers on MCP server integrations and composable, durable workflows, while deepagents provides planning, delegation, file access, and context management.
  • fast-agent: fast-agent emphasizes declarative multimodal agents and MCP workflows, while deepagents bundles planning, delegation, file access, and context management.
  • autogen: AutoGen focuses on multi-agent orchestration and is in maintenance mode, while deepagents offers a ready-made harness with planning, file access, and context management.
ProjectLanguageLicenseStarsStatus
deepagentsPythonMIT29.9kNot checked yet
mcp-agentPythonApache-2.08.6kQuiet
fast-agentPythonApache-2.03.9kActive
autogenPythonCC-BY-4.061.3kQuiet

Considerations

  • The model must support tool calling. Model choice is flexible, but provider or local-model setup is still required.
  • The agent can do anything its tools permit. Enforce security boundaries in the tools and sandbox rather than relying on the model to self-restrict.
  • Deep Agents provides more built-in behavior than a minimal agent harness. Use LangChain's create_agent for a lighter option, or LangGraph directly when you need to define a custom graph.
  • The project is designed for production deployments on LangGraph, but tracing, evaluation, and monitoring are paired with LangSmith rather than described as automatic defaults.

Found this useful?

Share it with someone who would like deepagents.

Comparisons