Advanced12 min
Agent Server: Architecture & Deployment
Agent Server (formerly LangGraph Platform) is the production runtime for LangGraph agents — with Assistants, Threads, Runs, Cron jobs, and three deployment modes from single host to distributed.
Quick Reference
- →Agent Server = managed runtime for LangGraph agents with persistence, streaming, and HITL built in
- →Core concepts: Assistants (agent configs), Threads (conversations), Runs (executions), Crons (scheduled)
- →Container architecture: stateless API servers + queue workers, PostgreSQL + Redis
- →Three deployment modes: single host, split API/queue, distributed runtime
- →Three durability modes: exit (fastest), async (balanced), sync (most durable)
- →Framework-agnostic: supports LangGraph, Deep Agents, Google ADK, Strands
Core Concepts
| Concept | What It Is | Analogy |
|---|---|---|
| Assistant | A versioned agent configuration (model, tools, prompt, graph) | A deployed application |
| Thread | A conversation with checkpointed state | A user session |
| Run | A single execution within a thread | A request/response cycle |
| Cron | A scheduled recurring run (with or without a thread) | A cron job |
| Store | Cross-thread persistent key-value storage | A user preferences database |
These concepts map to REST endpoints: POST /assistants, POST /threads, POST /threads/{id}/runs, POST /crons. The SDK wraps these into a clean Python/JS API. Every run is automatically checkpointed, so you can resume interrupted runs, time-travel to previous states, and stream intermediate results.