Deep Agents Orchestration Skill: Subagents, Todo Lists and Approvals

Teaches Claude how to add subagents, task planning and human approval to a LangChain Deep Agents project, in Python or TypeScript.

Installs: 16.4K; on skills.sh
Stars: 1.3K; on GitHub
Updated: Oct 2026; last repo commit

Deep Agents Orchestration is a LangChain guide that Claude reads when you build with Deep Agents and need subagents, a todo list or approval steps. It gives Claude working Python and TypeScript examples for each, so the code it writes follows the library's real patterns.

It is a reference, not a tool that runs anything. It bundles only SKILL.md and no scripts. You still need the deepagents package and a model provider key in your own project.

Good for

Developers building agents with the deepagents package
Teams that need a person to approve risky agent actions such as database writes or deploys
Claude Code users who want fewer mistakes with subagents, checkpointers and thread IDs

Install deep-agents-orchestration

Claude Code

Copy this prompt.

Install the deep-agents-orchestration skill in Claude Code. Run this command:

/plugin marketplace add langchain-ai/langchain-skills
/plugin install langchain-skills@langchain-skills

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

deep-agents-orchestration stands alone: its SKILL.md does not call other skills. The plugin route installs all 21 skills in langchain-skills (plugin.json points at config/skills/ and declares no hooks or MCP servers). The README marks the project as early development. Pick one route, not both.

From the project README. Source.

What's in the skill

When it loads

INVOKE THIS SKILL when using subagents, task planning, or human approval in Deep Agents. Covers SubAgentMiddleware, TodoList for planning, and HITL interrupts.

Sections in SKILL.md

Subagents (Task Delegation)
TodoList (Task Planning)
Human-in-the-Loop (Approval Workflows)
What Agents CAN Configure
What Agents CANNOT Configure

Key facts

Author: LangChain
License: MIT
Installs on skills.sh: 16,437 (Nov 2, 2026)
GitHub stars: 1,320 for the whole langchain-ai/langchain-skills repo (Nov 2, 2026)
Last commit: Oct 29, 2026
What is inside: 1 file: SKILL.md. No scripts or other files.
Starts: Claude loads it when a task matches its description, which begins "INVOKE THIS SKILL"
Repo status: The README marks the project as early development; APIs and skill content may change

What it includes

Delegate work to a fresh specialized agent that returns a report
The main agent calls a task tool, a new subagent works alone, and a final report comes back. A general-purpose subagent is always available. SKILL.md shows how to add a custom "researcher" with its own tools.
Try: “Add a researcher subagent with a paper search tool to my Deep Agents app.”
Subagents (task tool)
Plan and track multi-step work with pending, in_progress and completed items
Included by default. SKILL.md shows how to read the todo list from the final state and notes you need a thread_id to keep todos between calls.
Todo list (write_todos)
Pause before chosen tools run so a person can approve, edit or reject
You list the tools that need approval, such as write_file or execute_sql. A checkpointer is required, and the agent pauses between invoke calls until you resume it with approve, edit or reject.
Try: “Make my agent ask for approval before it runs execute_sql.”
Human-in-the-loop (interrupt_on)
Set interrupt_on on a single subagent such as a deployer
Shows a code-deployer subagent where only deploy_to_prod needs approval.
Subagent approvals
Tool names and the approve/edit/reject protocol are fixed
You can change subagent names, tools, models, prompts and which tools need approval. You cannot rename the task and write_todos tools or make subagents remember earlier calls.
What you can and cannot configure
Stateless subagents, missing skills, checkpointer and thread_id errors
Subagents start fresh each call, so give full instructions. Custom subagents do not inherit the main agent's skills. Interrupts need a checkpointer and a consistent thread_id to resume.
Common fixes

From the official docs

Questions

What does the deep-agents-orchestration skill do?

It gives Claude LangChain's own guidance and code examples for subagents, todo planning and human approval in Deep Agents. Claude uses it when you build or fix those parts of a Deep Agents app.

How do I install it?

Use the install box at the top of this page. The README of langchain-skills also documents a Claude Code plugin and an install script, and says npx skills works for any agent that supports the Agent Skills specification.

Does it need an API key?

The skill itself needs none. The agents it helps you write need a model provider key in your own project, and the README tells you to export OPENAI_API_KEY or ANTHROPIC_API_KEY.

Is it safe to install?

It is one text file with examples and bundles no scripts, so it adds instructions for Claude and does not run anything itself. Review any code Claude writes from it before you run it, especially approval settings, because the examples show write_file and SQL tools.

Does it work in Codex or Cursor?

The README says langchain-skills installs through npx skills for any agent that supports the Agent Skills specification, including Claude Code, Cursor and Windsurf, and shows a Codex command. It does not describe client-specific testing of this skill.

What is the difference from deep-agents-core?

The README describes deep-agents-core as agent architecture, harness setup and the SKILL.md format, and deep-agents-orchestration as subagents, task planning and human-in-the-loop. Use core for the basic agent and orchestration for delegation and approvals.

Information

updated Nov 2, 2026

Repository: langchain-ai/langchain-skills
Folder: config/skills/deep-agents-orchestration

langchain-ai/langchain-skills skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 3

github.com · github.com
github.com · github.com
github.com · github.com