Summary
Orchestration is the layer above individual agents: routing tasks, managing shared state, handling failure, aggregating results. Frameworks (LangGraph, AutoGen, CrewAI, Temporal) formalize the patterns. The right level of orchestration is task-dependent; too much is a common failure mode.
Objectives
- 01State what an orchestrator does.
- 02Distinguish state machines from event-driven orchestration.
- 03Recognize over-orchestration.
The Lesson
The layer
Orchestration handles: which agent runs when, what state they share, how failures are retried, how results merge. It sits between application code and individual agent calls. In simple systems it is a for-loop; in production it is a framework.
Over-orchestration
A 10-node LangGraph for a task a single prompt would handle is a red flag. Reach for orchestration when: multiple parallel agents, retriable steps with side effects, long-running workflows crossing failure boundaries.
Key Ideas
- Orchestration = flow control above agent calls.
- For simple tasks, a for-loop is orchestration.