85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Tool explainer

What is LangChain Deep Agents?

Deep Agents is LangChain's open-source, MIT-licensed agent harness, built on LangGraph, that gives any tool-calling model planning, a filesystem, subagents, context summarisation and memory out of the box for long-horizon tasks.

Checked

A plain tool-calling loop works for short tasks and struggles on long ones: context fills up, the agent loses its plan, and one thread tries to do everything. A pattern that holds up better combines a planning step, a filesystem to offload work to, subagents with their own context, and summarisation when the window gets full. Deep Agents packages that pattern as an open-source library that works with any model that supports tool calling.

It is the third layer of LangChain's stack. LangGraph is the runtime, LangChain's create_agent is the agent layer, and Deep Agents sits on top as the ready harness, in Python and as deepagents.js.

How it works

create_deep_agent takes a model string, a system prompt and any extra tools, and returns a LangGraph agent with streaming, persistence and checkpointing. The defaults are tuned for multi-step work, and any part can be replaced.

  • Filesystem

    Read, write, edit, search and delete over pluggable local, sandboxed or remote backends, so large tool outputs go to disk instead of the context window.

  • Subagents

    Delegated tasks run in isolated context windows and return a result, keeping the main thread short.

  • Context management and memory

    Long threads are summarised, and pluggable state and store backends carry memory across sessions.

  • Skills, tools and approvals

    Skills load on demand, custom functions and MCP servers add tools, and human-in-the-loop lets a person approve, edit or reject a tool call before it runs.

Version and status, checked 1 November 2026

  • Versions

    Python deepagents 0.7.32 (23 October 2026, Python 3.11 or later) and npm deepagents 1.16.2 (30 October 2026). The Python package is still pre-1.0.

  • Adoption

    30,841 GitHub stars and 1.85M PyPI downloads in the week to 1 November 2026.

  • v0.7 (29 July 2026)

    LangChain reported 65% fewer base input tokens at comparable performance: the default system prompt was removed, built-in tool descriptions were trimmed by 43%, and to-do lists became opt-in.

  • Hosting

    Managed Deep Agents entered public beta on 7 August 2026 as a hosted runtime on LangSmith, deployed with one command, limited at launch to LangSmith Cloud in the United States region. A terminal coding agent, Deep Agents Code, ships alongside.

When to pick it, and when not to

  • Pick it for

    Long-horizon research, coding and document agents where the team wants a ready harness but must keep the model choice open, or mix frontier, open-weight and local models.

  • Pick a lab SDK instead when

    The team has settled on one model family and wants that lab's own tested harness, such as the Claude Agent SDK on Claude.

  • Pick plain LangGraph instead when

    The flow is mostly fixed steps with a few model decisions. A hand-built graph is easier to test than a general harness.

Gotchas from production

  • Trust the LLM, then sandbox

    The project's security model trusts the model, so boundaries must be enforced at the tool or sandbox level, not in the prompt.

  • A delete tool by default

    Since v0.7 the default filesystem tools include delete. Agents that should not remove files need a tool allowlist.

  • Breaking changes before 1.0

    v0.7 turned to-do lists off by default and removed backend factories. Pin the version and re-run evals on each upgrade.

  • Token savings vary by model

    LangChain's own v0.7 results showed total token use and cost moving differently per model, with one model's cost rising on autonomous tasks. Measure on the model actually deployed.

  • Two version lines

    The Python package is on 0.7 while the JavaScript package is on 1.16, so release notes and examples need matching to the right language.

How Motiontic approaches it

Motiontic's engineers are vendor-neutral. They run a short spike on the client's real task before committing to a framework or store, keep prompts, tool definitions and evals outside the framework so a later switch is a refactor rather than a rewrite, and hand the eval suite to the client's team so the next tooling decision can be measured rather than argued.

Deep Agents and its nearest alternatives, checked 1 November 2026

Deep Agents and its nearest alternatives, checked 1 November 2026
Deep AgentsClaude Agent SDKLangGraph
What it isReady agent harness on LangGraphClaude Code's agent loop as a libraryLow-level runtime for graph-shaped agents
ModelsAny tool-calling modelClaudeAny model the team connects
Built inPlanning, filesystem, subagents, summarisation, memoryFiles, shell, web search, subagents, hooks, permissionsGraph, state, checkpoints, human-in-the-loop
LicenceMITAnthropic Commercial Terms of ServiceMIT
Latest version0.7.32 Python (23 Oct 2026), 1.16.2 JS0.2.176 Python (31 Oct 2026)1.2.19 Python (30 Oct 2026)
Hosted optionManaged Deep Agents (beta, LangSmith)Claude Managed Agents (beta)LangChain and Google managed runtimes
PyPI downloads, week to 1 Nov 20261.85M8.09M13.60M

Common questions

What is the difference between Deep Agents and LangGraph?

LangGraph is the runtime for building agents as graphs. Deep Agents is a finished harness built on LangGraph, with planning, a filesystem, subagents and memory already wired in.

Is Deep Agents an alternative to the Claude Agent SDK?

It covers similar ground: both are harnesses for long-running agents with files and subagents. Deep Agents is MIT-licensed and model-neutral; the Claude Agent SDK runs Claude Code's own loop on Claude models.

Which models does Deep Agents support?

Any model that supports tool calling. The README lists OpenAI, Anthropic, Google, Ollama, vLLM and llama.cpp among the options.

Can Deep Agents be hosted?

Yes. Managed Deep Agents, in public beta since 7 August 2026, runs Deep Agents on LangSmith. Because it is an open-source library built on LangGraph, it can also run in a team's own infrastructure.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.