feat: subagent turn count, simple agent loop

[wendytang]

• allow a maximum of 5 subagent loops
• kept fairly small number of loops for now to ensure subagent tasks complete.
• interesting problem to think about: depending on complexity of subagent tasks, size up or down # agent loop permitted

[lifeizhou-ap]

Hi @wendytang ,

While reviewing this PR, I feel the subagent has part of the agent(agent.rs) reply() logic but not all. in the agent we have max turn logic, final result and retry logic. These are the logic that we would also to have in the subagent. Just wondering whether it is possible to refactor the code to reuse the agent reply? so that we don't have to duplicate the logic every time when we have a new feature in agent

[DOsinga]

independent of this, is there a plan to unify this subagent loop with the main agent loop? already I feel l like we're starting to see code drift between the two approaches

[wendytang]

I agree that we should unify subagent -> simply use agent reply().

Dug into this a bit, and it boils down to the reply method's return type. It reaches the upper limits of my rust knowledge for now, but I do believe this is possible.

The error occurs because you're trying to call child_agent.reply() within the Agent::reply() method, but the future returned by reply() is not Send. This is a common Rust async issue related to thread safety.
Looking at the reply() method signature on line 876:

pub async fn reply(
    &self,
    messages: &[Message],
    session: Option<SessionConfig>,
    cancel_token: Option<CancellationToken>,
) -> Result<BoxStream<'_, Result<AgentEvent>>> {