Deep Agent — Pydantic AI

A Pydantic AI re-implementation of the LangChain Deep Agents pattern, which itself was reverse-engineered from Claude Code's architecture.

What is this?

LangChain's Deep Agents project identified four architectural pillars that make Claude Code effective at complex, multi-step tasks. This project recreates those pillars using Pydantic AI instead of LangChain/LangGraph:

Installation

pip install pydantic-ai-deepagents

Quickstart

import asyncio
from pydantic_ai_deepagents import create_deep_agent

async def main():
    agent, deps = create_deep_agent(
        system_prompt="You are a research assistant.",
    )

    result = await agent.run(
        "Research X and write your findings to notes.md",
        deps=deps,
    )

    print(result.output)

    # Check what files the agent created
    for path, entry in deps.files.items():
        print(f"{path}: {len(entry.content)} chars")

asyncio.run(main())

Customisation

Custom model

agent, deps = create_deep_agent(
    model="openai:gpt-4o",  # or any pydantic-ai model string
)

Custom tools

from pydantic_ai import RunContext
from pydantic_ai_deepagents import DeepAgentDeps

async def search_web(ctx: RunContext[DeepAgentDeps], query: str) -> str:
    """Search the web for information."""
    # your implementation
    return "results..."

agent, deps = create_deep_agent(tools=[search_web])

Disk-backed workspace with shell

from pathlib import Path

agent, deps = create_deep_agent(
    workspace=Path("./my_workspace"),
    enable_shell=True,
)

Sub-agent depth control

agent, deps = create_deep_agent(
    max_sub_agent_depth=3,  # allow 3 levels of nesting
    enable_subagents=True,
)

Skills

Skills follow the open Agent Skills specification with progressive disclosure:

1. Discovery — at startup, only skill names + descriptions go into the system prompt
2. Activation — the agent calls read_skill("skill-name") to load full instructions on demand
3. Execution — the agent follows the instructions, optionally loading resource files

Loading skills

# From directories
agent, deps = create_deep_agent(skills=["./skills/"])

# Inline
from pydantic_ai_deepagents import Skill
agent, deps = create_deep_agent(inline_skills=[Skill(
    name="my-skill", description="...", instructions="...", source_path="<inline>",
)])

# From raw SKILL.md text
from pydantic_ai_deepagents import load_skill_from_text
skill = load_skill_from_text("my-skill", open("SKILL.md").read())
agent, deps = create_deep_agent(inline_skills=[skill])

Example skills included

The example_skills/ directory contains three ready-to-use skills:

• web-research — structured research workflow with synthesis
• code-review — multi-pass review (correctness, security, maintainability, performance)
• data-analysis — data profiling, cleaning, analysis, and reporting

Credits

• langchain-ai/deepagents — the original Deep Agents implementation
• Claude Code — the architecture that inspired it all
• Pydantic AI — the agent framework used here