LangGraph: Graph-Based State Machines

The loop IS the graph — the largest architectural divergence in the roster. Every other harness hides its loop in an implicit while-true. LangGraph makes the loop a declared, visible, testable, editable graph you draw before you run it. Nodes are functions (take state, do work, return a state update). Edges are transitions, with conditional edges expressing branching in code. Module 1 (Loop): 5/5 — the reference. This is the framework for when the process is the product.

Super-step checkpoints are the Module 8 reference at 5/5 — the finest-grained state model in the roster. State is serialized at every node boundary. The granularity is the STEP, not the session. A crash resumes from the last completed node. For multi-day, multi-human workflows this is the difference between 'resume from where you were' and 'resume from approximately where you were.' Every checkpoint is serializable, replayable, forkable, inspectable.

interrupt() is the cleanest HITL primitive in the roster — structural, not behavioral. The edge from propose to execute goes through an interrupt node. The model cannot route around the approval because the model does not control the edges — the graph definition does. A prompt-injected agent cannot skip the approval because the approval is a node in the graph topology, not a request in the prompt. Same principle as NemoClaw (DD-09): enforcement outside the agent's reach.

Subgraphs provide declared multi-agent coordination — the C4 E09 pole. A multi-agent system is a graph of subgraphs: each subagent is a node, edges between subagents are declared transitions in code. LangGraph is the declared-coordination pole; CrewAI (DD-12) is the emergent-coordination pole. The trade-off is the same as the single-agent case: declared edges are auditable and testable but rigid; emergent handoffs are flexible but opaque. The contrast defines the multi-agent axis the way NemoClaw-vs-Tau defines the governance axis.