Foundation framework for building multi-step agents with built-in planning, memory, and skill delegation. Provides six core middleware options: task planning, filesystem context management, subagent delegation, persistent memory, human approval workflows, and on-demand skill loading Includes three always-present built-in tools: write_todos for task tracking, filesystem operations ( ls , read_file , write_file , edit_file , glob , grep ), and task for spawning specialized subagents Supports...

1 file — installable on its own

Pluggable memory and file backends for Deep Agents with ephemeral, persistent, and hybrid routing options. Four backend types: StateBackend (thread-scoped, ephemeral), StoreBackend (cross-session persistent), FilesystemBackend (real disk access for local dev), and CompositeBackend (route different paths to different backends) FilesystemMiddleware provides six file operation tools: ls , read_file , write_file , edit_file , glob , grep CompositeBackend uses longest-prefix matching to route...

1 file — installable on its own

Orchestrate subagents, plan multi-step tasks, and require human approval for sensitive operations. Delegate work to specialized subagents via the task tool; custom subagents support isolated tool sets and system prompts, while the default "general-purpose" subagent inherits main agent configuration Plan and track complex workflows with write_todos , organizing tasks across pending, in-progress, and completed states; requires a thread_id for persistence across invocations Implement...

1 file — installable on its own

LangChain ecosystem package versions, dependencies, and installation guidance for Python and TypeScript. Start all new projects on LangChain 1.0 LTS; version 0.3 is legacy maintenance-only. Always install langchain-core explicitly alongside other packages. Choose one orchestration layer: LangGraph for fine-grained graph control, or Deep Agents for batteries-included planning and memory. Pin langchain-community conservatively to exact minor versions (e.g., >=0.4.0,<0.5.0 ) since it does not...

1 file — installable on its own

Build production LangChain agents with create_agent(), tools, and middleware patterns. Use create_agent() with model, tools list, and system prompt; configure state persistence with checkpointer and thread_id for conversation memory across invocations Define tools via @tool decorator (Python) or tool() function (TypeScript) with clear descriptions so agents know when to call them Add middleware like HumanInTheLoopMiddleware for approval workflows, custom error handling, and human-in-the-loop...

1 file — installable on its own

Human-in-the-loop approval, custom middleware, and structured output patterns for LangChain agents. HumanInTheLoopMiddleware pauses execution before dangerous tool calls, allowing humans to approve, edit arguments, or reject with feedback Per-tool interrupt policies let you configure different approval rules based on risk level; requires a checkpointer and thread_id for state persistence Command resume pattern continues execution after human decisions, with support for editing tool arguments...

1 file — installable on its own

Complete RAG pipeline for document ingestion, embedding, retrieval, and LLM-powered response generation. Supports multiple document loaders (PDF, web pages, directories) and persistent vector stores (Chroma, FAISS, Pinecone) with configurable chunk size and overlap for optimal context preservation Includes similarity search, MMR (Maximal Marginal Relevance) retrieval, and metadata filtering to balance relevance and diversity in results Works with OpenAI embeddings and integrates seamlessly...

1 file — installable on its own

Directed graph framework for building stateful, multi-step agent workflows with fine-grained control. StateGraph with typed state schemas, reducers for accumulating lists/values, and nodes that return partial state updates Static edges for fixed flow, conditional edges for branching, and Command for combining state updates with dynamic routing Send API for fan-out parallelism to worker nodes with result aggregation via reducers Invoke for single execution and stream modes (values, updates,...

1 file — installable on its own

Pause graph execution for human review, approval, or validation, then resume with their input. Requires three components: a checkpointer (InMemorySaver or PostgresSaver), a thread ID in config, and JSON-serializable interrupt payloads interrupt(value) pauses and surfaces data; Command(resume=value) resumes and returns that value to the paused node All code before interrupt() re-executes on resume, so side effects must be idempotent (use upsert, not insert) Supports approval workflows,...

1 file — installable on its own

Durable graph execution with thread-scoped checkpoints, state history, and cross-thread long-term memory. Three checkpointer options: InMemorySaver for testing, SqliteSaver for local development, PostgresSaver for production; always pass thread_id in config to enable persistence Browse and replay from past checkpoints using get_state_history() , fork execution by updating state at a past point, or manually modify state before resuming Store API provides cross-thread memory for user...

1 file — installable on its own