Fan out the agents. Cap how many run.

from graphingest import graph, deploy, ThrottlePolicy, ConcurrencyPolicy
from graphingest.langgraph import agent_node, AgentConfig

def build_researcher(config: AgentConfig):
    from langgraph.graph import StateGraph, END
    from typing import TypedDict

    class State(TypedDict):
        messages: list

    def reason(state: State) -> dict:
        question = state["messages"][-1]["content"]
        # Replace this with ChatOpenAI(model=config.model).invoke(...)
        answer = f"[{config.model}] notes on: {question}"
        return {"messages": [{"role": "assistant", "content": answer}]}

    builder = StateGraph(State)
    builder.add_node("reason", reason)
    builder.set_entry_point("reason")
    builder.add_edge("reason", END)
    return builder.compile()

researcher = agent_node(
    name="researcher",
    graph_builder=build_researcher,
    config=AgentConfig(model="gpt-4o-mini", max_iterations=6, stream_steps=False),
)

@graph(
    name="research-fanout",
    timeout_seconds=3600,
    throttle=ThrottlePolicy(limit=20, period_seconds=60),
    concurrency=ConcurrencyPolicy(limit=5, wait_timeout_seconds=180),
)
def research_fanout(queries: list[str]):
    width = 5  # agents in flight inside this run
    results = []
    for start in range(0, len(queries), width):
        results.extend(researcher.map(queries[start : start + width]))
    return results

if __name__ == "__main__":
    deploy()