Built-in tool that enables an agent to transfer control to a sub-agent.

When an LLM agent has sub-agents configured, transfer_to_agent tools are automatically generated — one per sub-agent. When the LLM calls this tool, it produces an event with actions.transfer_to_agent set, which the runner picks up to hand off execution to the target agent.

Differences from Python ADK

Python ADK creates a single transfer_to_agent tool with an enum constraint on the agent_name parameter. Elixir creates one tool per target agent (named transfer_to_agent_<name>) with the target hardcoded in the closure.

Both approaches prevent hallucinated agent names:

• Python: enum constraint restricts valid values
• Elixir: separate tools mean the LLM picks the right tool by name, no parameter to hallucinate

The per-agent approach was chosen for Elixir because:

1. It eliminates a class of errors (wrong agent_name parameter)
2. Tool descriptions per-agent give the LLM better context for selection
3. It works naturally with Elixir's pattern matching

How it works

1. The parent agent's tool list is augmented with one transfer_to_agent_* tool per sub-agent
2. When the LLM invokes the tool, it returns a transfer event
3. The parent agent's run loop detects the transfer and delegates to the target sub-agent
4. The sub-agent runs and its events are returned