Deep Agents

• https://github.com/langchain-ai/deepagents

• https://docs.langchain.com/oss/python/deepagents/cli/overview

• It's a new project in the Langchain umbrella

• It's basically a more rich agent with more capabilities

• Deep Agents solve the problem of an increasingly context. Because it delegates tasks and offload unnecessary context

• It's an Agent Harness (and agent general purpose client), not a framework like langchain

• Those capabilities/techniques mainly involve

• Planning Tool: e.g., checklists
• Sub Agents: sub instances of agents for specialized tasks
• File System
• System Prompt

Configuration

• ~/.deepagents/config.toml

Install

# install globally
uv tool install 'deepagents-cli[ollama,groq]'

Usage

# Run it (uses a default agent)
deepagents

# Prompt
git diff | deepagents --skill code-review -n 'summarize changes' # piped content appear first
deepagents -n "Generate a .gitignore for Python" -q > .gitignore

# Run your own agent
# the state (long term memory) for this agent will be stored (at ~/.deepagents/<myagent>/) and you can reopen it later on
deepagents --agent myagent

# List all agents created
deepagents agents list # default agent first

Skills

• Deep agents have a skills framework
• Workflow
• Skills are loaded to the to the agent state on the session start
• Then, on each LLM call, the system prompt is updated with the skills metadata.

• Based on that, the LLM can request the read of the full skill when needed (progressive disclosure)

• Skills at ~/.agents/skills/ are read by deepagents

• It has the built-in skill skill-creator

npx skills add "remotion-dev/skills"

• You can ask deepagents which skills he has to verify