---
title: "Deep Agents — Autonomous Multi-Step AI Agents"
tags: [ai-agents-autonomy-planning-file-operations-delegation-sandbox]
created: 2026-04-15
updated: 2026-04-30
type: concept
sources:
  - "https://docs.langchain.com/oss/python/deepagents/overview"
  - "https://github.com/langchain-ai/deepagents"
---

# Deep Agents — Autonomous Multi-Step AI Agents

## Definition

Deep agents are autonomous AI agents that combine multiple architectural patterns to handle complex, multi-step tasks with minimal human intervention. They feature:

1. **Planning** — Breaking complex tasks into sub-goals
2. **Progress tracking** — Monitoring completion status
3. **File operations** — Reading, writing, and modifying files
4. **Task delegation** — Spawning sub-agents for specialized work
5. **Sandboxed code execution** — Running generated code safely

## Pydantic AI's Approach

Pydantic AI defines Deep Agents as the highest level (level 6) in its multi-agent complexity hierarchy:

1. Single agent workflows
2. Agent delegation (agents calling other agents via tools)
3. Programmatic agent hand-off
4. Graph-based control flow
5. Deep Agents (autonomous, planning, file ops, delegation, sandbox)

## Implementation Patterns

### Planning and Progress Tracking
Deep agents break complex tasks into steps and track their progress, giving users visibility into the agent's reasoning and current state.

### File Operations
Unlike simple tool-calling agents, deep agents can read and write files, enabling them to:
- Modify codebases
- Create reports and documentation
- Manage configuration files

### Task Delegation
Deep agents can spawn sub-agents for specialized work, similar to how human engineers delegate to specialists.

### Sandboxed Execution
Deep agents require secure code execution environments. Options include:
- **Monty** — Minimal, secure Python interpreter (0.004ms start)
- **Docker** — Full isolation with higher overhead
- **Sandbox services** (Modal, E2B, Daytona) — Remote execution

## Community Implementations

- `pydantic-deep` by Vstorm — Opinionated package bringing deep agent patterns together

## Related

- [[concepts/pydantic-ai]] — Framework supporting deep agents
- [[concepts/harness-engineering]] — Environment design for autonomous agents
- [[concepts/code-mode]] — Code execution for deep agents
- [[concepts/monty-sandbox]] — Secure execution environment
- [[concepts/agent-architecture-decomposition]] — Model/Runtime/Harness three-layer framework
- [[entities/harrison-chase]] — LangChain CEO, Deep Agents framework originator
- [[entities/nvidia-openshell]] — Open Runtime reference implementation
- [[samuel-colvin]] — Pydantic AI creator

## Harrison Chase's Framework: Deep Agents as Open Harness

LangChain's Deep Agents is the reference implementation of an **Open Harness** in Harrison Chase's three-layer model. In this framework:

- **Model layer** → Any LLM (Claude, GPT, Nemotron) — model-agnostic
- **Runtime layer** → The execution environment (bash, Python REPL, Docker sandbox) — determines native tool-use interface
- **Harness layer** → Deep Agents provides planning, sub-agent spawning, memory management, tool routing

Deep Agents maps to the Harness layer, connecting the model to the runtime. Its design philosophy aligns with the "[[concepts/harness-engineering]]" principle: **Agent = Model + Harness**, where the harness is everything that wraps a raw model to turn it into a productive work engine.