Orchestration
The control logic that decides what an agent does next: which tool to call, which agent runs, when to branch, and when to stop.
Orchestration is the glue logic of an agent: the part that decides what happens next. Should the agent call a tool, hand off to another agent, branch based on a result, retry, or finish? That decision-making is what separates a real agent system from a single prompt.
Frameworks express orchestration in different shapes. LangGraph makes it an explicit graph of nodes and edges, so the control flow is visible and inspectable. CrewAI hides it behind roles and tasks, which is simpler but less transparent. AutoGen drives it through agent conversation. Each shape trades visibility against ease of use.
How much orchestration you need scales with complexity. A single agent with three tools needs almost none; the model decides each step and the loop just runs. A workflow with branching paths, multiple agents, retries, and approval gates needs real orchestration, and that is where a framework's control model becomes the deciding factor. Pick the orchestration style that matches how complex your real workflow will get, not the simplest demo.