LangChain

Wires up persistence layers for a Deep Agent — choose between ephemeral session state, durable storage, or filesystem access. Routes automatically based on what the agent needs.

Configures subagents to hand off work, plans tasks in a shared list, and pauses for human sign-off before critical steps—keeping complex workflows from spinning out of control.

Answers whether you need LangChain, LangGraph, Deep Agents, or a hybrid — and why. Run this first, before building any agent logic or tool integration.

Handles environment setup, package version matching, and dependency conflicts for LangChain, LangGraph, and LangSmith. Catches breaking changes and suggests pin strategies.

Scaffolds a working LangChain agent with tool definitions, middleware for human approval steps, and error handlers so the agent won't break production on its first mistake.

Reference for the core Deep Agents API: how to spin up an agent, wire it to skills, set memory, configure execution loops, and deploy. Covers the harness architecture end-to-end.

Guides you through StateGraph setup, node logic, branching, and streaming so your agent chains steps predictably. Includes error-handling patterns and state-schema design.

Configures checkpointers, thread scoping, and conversation history so your graph remembers past runs and can travel backward through state without losing context or crashing on replay.

Takes a table or list, splits each row into an independent task, runs them all at once, then stitches the results back together into one output.

Scaffolds a working retrieval-augmented generation system: loads documents, chunks them smartly, embeds vectors, and connects a vector store so Claude can ground answers in your data.

Scaffolds a new LangGraph project, walks through dev mode, handles langgraph.json config, and deploys to the cloud with a single command chain.

Explains how to pause a LangGraph agent mid-run, collect human feedback or approval, and resume—covering interrupts, Commands, and multi-tier error handling patterns.

Integrates human approval gates, custom middleware hooks, and structured-output validation into LangChain agents—so dangerous or high-stakes tool calls pause for review before executing.

Deploys a coordinated team of Claude agents that talk to each other via MCP servers, each with their own tools and skills. Handles threading, interrupts, and live streaming.