What should I use for multi-agent orchestration?

It depends on what you mean by “multi-agent orchestration”:

If you want a framework to build agent workflows

Good options:

• LangGraph — best if you want stateful, controllable agent graphs with branching, loops, retries, and human-in-the-loop steps.
• Microsoft AutoGen — good for agent-to-agent conversation patterns and experimental multi-agent collaboration.
• CrewAI — simpler, more opinionated, good for role-based agent teams and quick prototypes.
• Semantic Kernel — solid if you’re already in the Microsoft ecosystem and want tool/plugin orchestration with agents.

If you want production orchestration of tasks, not just agents

Use a workflow engine and treat agents as workers:

• Temporal — excellent for reliable production workflows, retries, long-running jobs, and observability.
• Apache Airflow — better for batch/data pipelines than interactive agents.
• Prefect / Dagster — good for data/ML workflows, less agent-specific than Temporal.

My general recommendation

• Prototype / research: AutoGen or CrewAI
• Serious agent workflow with control: LangGraph
• Production-grade orchestration: Temporal + your agent framework
• Data pipelines: Airflow / Prefect / Dagster

Rule of thumb

• Need conversation among agents → AutoGen
• Need explicit graphs and state machines → LangGraph
• Need simple team setup → CrewAI
• Need durable execution and retries → Temporal

If you tell me:

1. your use case,
2. your stack (Python/JS/.NET),
3. whether this is prototype or production,

I can recommend one specific option.