Orchestrator agent

An orchestrator agent is the agent in a multi-agent system that receives the overall request, breaks it into subtasks, delegates them to other agents, and assembles the final result.

The pattern. Anthropic’s orchestrator-workers pattern has a central model decide the subtasks at run time, because they cannot be known in advance, then delegate them to worker models and synthesize their results. That flexibility separates it from a fixed pipeline, where code decides the steps.

In frameworks.

• Google ADK: a coordinator LlmAgent lists specialists in sub_agents. It delegates either by LLM-driven transfer, which relies on clear descriptions of each sub-agent, or by wrapping specialists with AgentTool and calling them explicitly.
• CrewAI: in the hierarchical process, a manager agent, set with manager_llm or manager_agent, plans the work, assigns tasks according to each agent’s capabilities, and reviews the outputs. Tasks are not pre-assigned.
• OpenAI Agents SDK: a triage agent either hands off the conversation to a specialist, or calls the specialist with Agent.as_tool and keeps control.

Hand off or call. The choice decides who owns the conversation. A handoff transfers it, and by default the specialist sees the whole history. Calling a specialist as a tool keeps the orchestrator in charge and passes only structured input, without transferring the conversation. The orchestrator then stays responsible for the final answer.

Across organizations. An orchestrator can delegate to an agent run by another company by acting as an A2A client: it sends the subtask as a message, tracks the resulting task, and treats the remote agent’s artifacts as one more result to combine. It cannot see the remote agent’s internal steps, so it relies on task states, and on its own checks of what comes back.

Neighbouring terms. Sub-agents are the workers it delegates to. A planner produces the step list, and in many systems the orchestrator plays both roles.