Create Subagent natively using AgentDefinition vs creating an agent and wrapping it as a tool for the lead agent

[kargarisaac]

Hello,

I have been trying to build a multi agent system with subagents specialized in processing some files like csv and have been trying to use AgentDefinition, but I face some permission issues and I don't see how can I provide permissions for the subagent in the doc:

@dataclass
class AgentDefinition:
    description: str
    prompt: str
    tools: list[str] | None = None
    model: Literal["sonnet", "opus", "haiku", "inherit"] | None = None

And it was failing allot with no permission error types. At the end, I created agents and wrapped them in @tool and gave them to the lead agent. But I think in this way the conversation history between the lead and the subagent is gone each time the lead calls the subagent as it creates a new session, right? any solution or thought on this? am I doing sth wrong or is this the sdk limitation? what is the write way to do it?