Function agent_turns 

pub fn agent_turns(
    tools: impl Tool + 'static,
    client: Client,
    base_request: Request,
    history: Vec<Message>,
    history_budget: Option<usize>,
) -> AgentTurnsStream

Drive the LLM ↔ tool loop using non-streaming Request::complete calls, yielding one crate::types::CompleteResponse per LLM turn.

Like agent but at turn granularity instead of token granularity: each item in the stream is a complete LLM response for one turn. Intermediate turns (where the model called tools) are yielded before tool execution; the final turn (no tool calls) is the last item.

Tool calls are executed concurrently between turns, exactly as in agent().

Usage patterns

use agentix::{ToolBundle, Request, Provider};
use futures::StreamExt;

let client = reqwest::Client::new();
let request = Request::new(Provider::OpenAI, "sk-...");

// Just the final text:
let text = agentix::agent_turns(ToolBundle::default(), client.clone(), request.clone(), vec![], None)
    .last_content().await;
println!("{text}");

// Full response (with usage, tool_calls, etc.):
let response = agentix::agent_turns(ToolBundle::default(), client.clone(), request.clone(), vec![], None)
    .last_ok().await;

// With per-turn progress:
let mut stream = agentix::agent_turns(ToolBundle::default(), client, request, vec![], None);
while let Some(Ok(resp)) = stream.next().await {
    eprintln!("turn: {} tool calls", resp.tool_calls.len());
}