[BUG] Managed Agents aren't registered as tools

[strawgate]

When using Gemini, only registered tools will be called by the LLM. The logic for starting a new managed agent is hidden in the call tool code and the managed agents aren't themselves exposed as tools. This means that tool calling agents cannot start managed agents on gemini

This is not a problem for the code agent as it doesn't directly make tool calls in the same way

[Mandark-droid]

Hey am using gemini with code agents and it's not effectively able to use the managed agents. Also will using another model via ollama or openai or transformers help solve the problem

[strawgate]

As a workaround, you can do something like this to create a tool for each agent and then place the tool in the calling agents tool list

def call_agent_tool(task: str) -> str:  # noqa: ARG001
    """Call a managed agent to perform a task.

    Args:
        task: The context the managed agent will receive when it is called. This must be enough information for the Agent
         to successfully complete the task. It should also include what you expect to receive from the agent as a response.
         More is more.

    Returns:
        The response from the managed agent.

    Example:
    >>> call_agent_tool("What is the capital of France?")
    "Paris"
    >>> call_agent_tool("What is 2+2?")
    "4"
    """
    return "Do Nothing"


def managed_agent_to_tool(agent_name: str, agent_description: str) -> Tool:
    new_tool = tool(call_agent_tool)
    new_tool.name = agent_name
    new_tool.description = agent_description
    new_tool.inputs = {"task": {"type": "string", "description": "The detailed task to be performed by the agent"}}
    new_tool.output_type = "string"
    return new_tool