Deep Agents (written “DeepAgents” in the post this page draws on) is an open-source package maintained by LangChain, which Harrison Chase in Agent Frameworks, Runtimes, and Harnesses- oh my! describes as its newest project and an increasingly popular one. He classifies it as an Agent Harness, distinguishing it from LangChain, an Agent Framework, and LangGraph, an Agent Runtime.
Capabilities
On the post’s account Deep Agents is higher level than an agent framework: it builds on top of LangChain and adds default prompts, opinionated handling for tool calls, tools for planning, and access to a filesystem, among other things. Chase’s summary is that it is more than a framework — it comes with batteries included.
The project also includes deepagents-cli, which LangChain calls its coding agent, and is available in Python and JavaScript. In Improving Deep Agents with harness engineering LangChain reports improving deepagents-cli on Terminal Bench 2.0 by changing only its harness — its system prompt, tools and Agent Middleware — while keeping the model fixed. The middleware described there inject the agent’s working directory and available tools at start, detect repeated edits to the same file, and remind the agent to verify its work against the task before exiting.
Adoption & Ecosystem
LangChain has also described Deep Agents as a “general purpose version of Claude Code”, referring to Claude Code. The post sets it beside the Claude Agent SDK, which it reads as Claude Code’s own step toward being an agent harness; besides that SDK, Chase says he did not think there were many other general-purpose agent harnesses at the time, while allowing that all the coding CLIs could be argued to be agent harnesses in a way.