Overview

The run_agent tool enables your agents to run other agent as a tool in a separate conversation and wait for their response. This is particularly useful to:

• Share behaviors factored in specialized custom agents across multiple higher-level agents. This lets you create higher-level agents from lower-level custom agents specialized at certain task for better maintainability.
• Delegate tasks in long-running agentic loops to sub-agents operating in separate conversation. Long agentic loop can rapidly exhaust the context of models. Delegating tasks to sub-agents allows running longer loops to achieve more complex tasks.

Configuration

The tool is configured as a regular tool in the agent builder and simply consists in selecting the agent to call.

You can provide context in the agent instructions about when to call the agent and what information to provide when doing so.

Accessing sub-agents conversations

When your agent calls another agent you will be able to access the query and response (as well as agent thoughts) in the action detail panel. A link is also present to introspect the sub-conversation created and the actions taken by the sub-agent there.

Limitations

We enforce a maximum recursion depth of 4 for calling sub-agents. Beyond this depth the action will return an error to the agent attempting it.

Citations are not currently natively handled across sub-agents call but we're working on it!