LangGraph

A framework (from the LangChain team) for building stateful, controllable LLM agent workflows as graphs. You model an agent as nodes (steps/tools/LLM calls) connected by edges (including conditional and cyclic ones), with shared state threaded through. Unlike a linear chain or a fully autonomous loop, the graph makes control flow explicit — which branches run, when to loop, where a human approves — and supports persistence so long-running agents can pause, resume, and recover.

Description

  • Graph model — nodes + edges (conditional, cyclic) for explicit agent control flow.
  • Shared state — typed state passed and updated across nodes.
  • Persistence / checkpointing — durable state so workflows resume after interruption.
  • Human-in-the-loop — pause for approval/edits mid-run.
  • Streaming & observability — stream steps/tokens; pairs with LangSmith tracing.
  • Multi-agent — compose supervisor/worker agent topologies.

Download or use

pip install langgraph        # or: npm i @langchain/langgraph

Reasoning for

LangGraph featured in the Tech To The Rescue platform’s agent layer — orchestrating multi-step agent workflows (alongside MCP for tool access) over the Next.js + Postgres stack. Its appeal in an enterprise-transition context is control and auditability: a tech-for-good org moving from ad-hoc automation to governed skills needs agent flows that are inspectable, resumable, and human-approvable, not opaque autonomous loops. The explicit graph + persistence model gives exactly that — which aligns with the engagement’s theme of turning shadow-IT automation into owned, auditable systems. It’s the orchestration counterpart to lighter skill frameworks like Agent Skills.

Alternatives considered

  • Plain Agent Skills / Claude Code orchestration — simpler and often enough; LangGraph adds explicit state-machine control for complex, resumable flows.
  • CrewAI / AutoGen — role/conversation-based multi-agent frameworks; more autonomous, less explicit control than LangGraph’s graph.
  • Temporal — durable workflow engine (general-purpose); more infra, not LLM-shaped out of the box.
  • Archon / Ruflo — agent harness/orchestration tools in the vault; different layer (dev harness vs in-app agent runtime).

Resources


Template: tool