deep-agent

A configurable, production-oriented deep agent built on LangGraph: an orchestrator that plans, delegates to specialized sub-agents (Planner, Researcher, Writer, Coder), manages memory, recovers from errors, and can pause for human input.

Capabilities

• Plan — Planner sub-agent decomposes the task; orchestrator tracks it as a TODO list.
• Act — tools for web search, a virtual file system, and code editing.
• Remember — SQLite checkpointing (resume by thread_id) + optional ChromaDB vector recall.
• Delegate — task (sequential) and task_batch (parallel fan-out) to sub-agents in isolated contexts.
• Recover — retry with backoff; failed tools return messages so the orchestrator reroutes.
• Ask — ask_human pauses via LangGraph interrupt() and resumes on Command(resume=...).

Install

uv sync          # or: pip install -e .
cp .env.example .env   # fill in OPENAI_API_KEY, TAVILY_API_KEY

Use

from deep_agent import create_deep_agent, AgentConfig

agent = create_deep_agent(AgentConfig())
result = agent.invoke(
    {"messages": [{"role": "user", "content": "Research MCP and write a report"}]},
    config={"configurable": {"thread_id": "task-001"}},
)

See examples/ for research, custom sub-agents, human-in-the-loop, and parallel fan-out.