Orchestration

agent_type selects how an agent executes.

Type	Behaviour
`llm` / `base`	A single agent, executed directly
`parallel`	Sub-agents run concurrently, results are combined
`sequential`	Sub-agents run one after another, each seeing the previous output
`loop`	Iterative execution until a stop condition is met

Sub-agents

Composite agents reference children through sub_agents. Children are themselves agents — they can carry their own tools, model and instruction, and can be composite in turn, which is how hierarchies are built.

{
  "agent_name": "Research pipeline",
  "agent_type": "sequential",
  "sub_agents": ["collector_agent_id", "analyst_agent_id", "writer_agent_id"]
}

Under Google ADK, sub-agents of one invocation share an execution tree. When reading traces, do not assume one identifier per agent — correlate on the agent name as well.

Get started

Concepts

Guides

Self-hosting

Ecosystem

Contributing

Sub-agents

​Sub-agents

Sub-agents