langgraph-architecture

Single graphs work for simpler workflows, while subgraphs enable modularity and reusability in complex multi-agent systems. Choose subgraphs when you need independent reasoning loops, cross-graph communication, or separate persistence boundaries for different agent responsibilities.

Persistence strategy depends on your recovery requirements and latency tolerance. LangGraph supports multiple checkpointer strategies; choose based on checkpoint granularity (per-node vs per-step), recovery scope, and whether you need resumable execution across runtime restarts.

LangGraph supports three primary patterns: supervisor (centralized routing), peer (distributed decision-making), and streaming (real-time output). Select based on whether you need hierarchical control, autonomous agent negotiation, or continuous result streaming across heterogeneous runtimes.

Define state schemas upfront to structure agent memory and routing decisions. State management determines graph partitioning, cross-graph communication constraints, and checkpoint recovery semantics—critical for robust multi-agent and streaming applications.

Static routing follows predefined paths and suits deterministic workflows. Dynamic routing adapts based on runtime state and agent decisions, enabling flexible multi-agent coordination but requiring careful state and persistence design.

Yes. LangGraph streaming modalities span heterogeneous runtimes through defined state schemas and checkpoint strategies. Design cross-graph communication and state synchronization to enable real-time streaming while maintaining persistence and recovery guarantees.