framework-selection

┌─────────────────────────────────────────┐ │ Deep Agents │ ← highest level: batteries included │ (planning, memory, skills, files) │ ├─────────────────────────────────────────┤ │ LangGraph │ ← orchestration: graphs, loops, state │ (nodes, edges, state, persistence) │ ├─────────────────────────────────────────┤ │ LangChain │ ← foundation: models, tools, chains │ (models, tools, prompts, RAG) │ └─────────────────────────────────────────┘ Picking a higher layer does not cut you off from lower layers — you can use LangGraph graphs inside Deep Agents, and LangChain primitives inside both. This skill should be loaded at the top of any project before selecting other skills or writing agent code. The framework you choose dictates which other skills to invoke next. Decision Guide Answer these questions in order: Question Yes → No → Does the task require breaking work into sub-tasks, managing files across a long session, persistent memory, or loading on-demand skills? Deep Agents ↓ Does the task require complex control flow — loops, dynamic branching, parallel workers, human-in-the-loop, or custom state? LangGraph ↓ Is this a single-purpose agent that takes input, runs tools, and returns a result? LangChain ( create_agent ) ↓ Is this a pure model call, chain, or retrieval pipeline with no agent loop? LangChain (LCEL / chain) — Framework Profiles LangChain — Use when the task is focused and self-contained Best for: Single-purpose agents that use a fixed set of tools RAG pipelines and document Q&A Model calls, prompt templates, output parsing Quick prototypes where agent logic is simple Not ideal when: The agent needs to plan across many steps State needs to persist across multiple sessions Control flow is conditional or iterative Skills to invoke next: langchain-models , langchain-rag , langchain-middleware LangGraph — Use when you need to own the control flow Best for: Agents with branching logic or loops (e.g. retry-until-correct, reflection) Multi-step workflows where different paths depend on intermediate results Human-in-the-loop approval at specific steps Parallel fan-out / fan-in (map-reduce patterns) Persistent state across invocations within a session Not ideal when: You want planning, file management, and subagent delegation handled for you (use Deep Agents instead) The workflow is straightforward enough for a simple agent Skills to invoke next: langgraph-fundamentals , langgraph-human-in-the-loop , langgraph-persistence Deep Agents — Use when the task is open-ended and multi-dimensional Best for: Long-running tasks that require breaking work into a todo list Agents that need to read, write, and manage files across a session Delegating subtasks to specialized subagents Loading domain-specific skills on demand Persistent memory that survives across multiple sessions Not ideal when: The task is simple enough for a single-purpose agent You need precise, hand-crafted control over every graph edge (use LangGraph directly) Middleware — built-in and extensible: Deep Agents ships with a built-in middleware layer out of the box — you configure it, you don't implement it. The following come pre-wired; you can also add your own on top: Middleware What it provides Always on? TodoListMiddleware write_todos tool — agent plans and tracks multi-step tasks ✓ FilesystemMiddleware ls , read_file , write_file , edit_file , glob , grep tools ✓ SubAgentMiddleware task tool — delegate work to named subagents ✓ SkillsMiddleware Load SKILL.md files on demand from a skills directory Opt-in MemoryMiddleware Long-term memory across sessions via a Store instance Opt-in HumanInTheLoopMiddleware Interrupt and request human approval before sensitive tool calls Opt-in Skills to invoke next: deep-agents-core , deep-agents-memory , deep-agents-orchestration Mixing Layers When to mix Scenario Recommended pattern Main agent needs planning + memory, but one subtask requires precise graph control Deep Agents orchestrator → LangGraph subagent Specialized pipeline (e.g. RAG, reflection loop) is called by a broader agent LangGraph graph wrapped as a tool or subagent High-level coordination but low-level graph for a specific domain Deep Agents + LangGraph compiled graph as a subagent How it works in practice A LangGraph compiled graph can be registered as a subagent inside Deep Agents. This means you can build a tightly-controlled LangGraph workflow (e.g. a retrieval-and-verify loop) and hand it off to the Deep Agents task tool as a named subagent — the Deep Agents orchestrator delegates to it without caring about its internal graph structure. LangChain tools, chains, and retrievers can be used freely inside both LangGraph nodes and Deep Agents tools — they are the shared building blocks at every level. Quick Reference LangChain LangGraph Deep Agents Control flow Fixed (tool loop) Custom (graph) Managed (middleware) Middleware layer Callbacks only ✗ None ✓ Explicit, configurable Planning ✗ Manual ✓ TodoListMiddleware File management ✗ Manual ✓ FilesystemMiddleware Persistent memory ✗ With checkpointer ✓ MemoryMiddleware Subagent delegation ✗ Manual ✓ SubAgentMiddleware On-demand skills ✗ ✗ ✓ SkillsMiddleware Human-in-the-loop ✗ Manual interrupt ✓ HumanInTheLoopMiddleware Custom graph edges ✗ ✓ Full control Limited Setup complexity Low Medium Low Flexibility Medium High Medium Middleware is a concept specific to LangChain (callbacks) and Deep Agents (explicit middleware layer). LangGraph has no middleware — you wire behavior directly into nodes and edges.