并发代理

本节探讨多个代理并发工作的使用场景，主要涵盖三种模式：

1. 单消息多处理器
   展示单个消息如何被订阅同一主题的多个代理同时处理。

2. 多消息多处理器
   演示如何根据主题将特定消息类型路由到专用代理。

3. 直接消息传递
   专注于代理之间以及运行时与代理之间的消息发送。

import asyncio
from dataclasses import dataclass

from autogen_core import (
    AgentId,
    ClosureAgent,
    ClosureContext,
    DefaultTopicId,
    MessageContext,
    RoutedAgent,
    SingleThreadedAgentRuntime,
    TopicId,
    TypeSubscription,
    default_subscription,
    message_handler,
    type_subscription,
)

@dataclass
class Task:
    task_id: str


@dataclass
class TaskResponse:
    task_id: str
    result: str

单消息多处理器

第一种模式展示单个消息如何被多个代理同时处理：

• 每个Processor代理使用default_subscription()装饰器订阅默认主题

• 当向默认主题发布消息时，所有注册代理将独立处理该消息

备注

下方我们使用default_subscription()装饰器订阅Processor，还有另一种完全不使用装饰器的订阅方式，如主题订阅与发布所示，这种方式可以让同一个代理类订阅不同主题。

@default_subscription
class Processor(RoutedAgent):
    @message_handler
    async def on_task(self, message: Task, ctx: MessageContext) -> None:
        print(f"{self._description} starting task {message.task_id}")
        await asyncio.sleep(2)  # 模拟工作
        print(f"{self._description} finished task {message.task_id}")

runtime = SingleThreadedAgentRuntime()

await Processor.register(runtime, "agent_1", lambda: Processor("Agent 1"))
await Processor.register(runtime, "agent_2", lambda: Processor("Agent 2"))

runtime.start()

await runtime.publish_message(Task(task_id="task-1"), topic_id=DefaultTopicId())

await runtime.stop_when_idle()

Agent 1 starting task task-1
Agent 2 starting task task-1
Agent 1 finished task task-1
Agent 2 finished task task-1

多消息多处理器

第二种模式演示将不同类型消息路由到特定处理器：

• UrgentProcessor订阅"urgent"紧急主题

• NormalProcessor订阅"normal"普通主题

我们使用type_subscription()装饰器让代理订阅特定主题类型。

TASK_RESULTS_TOPIC_TYPE = "task-results"
task_results_topic_id = TopicId(type=TASK_RESULTS_TOPIC_TYPE, source="default")


@type_subscription(topic_type="urgent")
class UrgentProcessor(RoutedAgent):
    @message_handler
    async def on_task(self, message: Task, ctx: MessageContext) -> None:
        print(f"Urgent processor starting task {message.task_id}")
        await asyncio.sleep(1)  # 模拟工作
        print(f"Urgent processor finished task {message.task_id}")

        task_response = TaskResponse(task_id=message.task_id, result="Results by Urgent Processor")
        await self.publish_message(task_response, topic_id=task_results_topic_id)


@type_subscription(topic_type="normal")
class NormalProcessor(RoutedAgent):
    @message_handler
    async def on_task(self, message: Task, ctx: MessageContext) -> None:
        print(f"Normal processor starting task {message.task_id}")
        await asyncio.sleep(3)  # 模拟工作
        print(f"Normal processor finished task {message.task_id}")

        task_response = TaskResponse(task_id=message.task_id, result="Results by Normal Processor")
        await self.publish_message(task_response, topic_id=task_results_topic_id)

在注册代理后，我们可以向"urgent"和"normal"主题发布消息：

runtime = SingleThreadedAgentRuntime()

await UrgentProcessor.register(runtime, "urgent_processor", lambda: UrgentProcessor("Urgent Processor"))
await NormalProcessor.register(runtime, "normal_processor", lambda: NormalProcessor("Normal Processor"))

runtime.start()

await runtime.publish_message(Task(task_id="normal-1"), topic_id=TopicId(type="normal", source="default"))
await runtime.publish_message(Task(task_id="urgent-1"), topic_id=TopicId(type="urgent", source="default"))

await runtime.stop_when_idle()

Normal processor starting task normal-1
Urgent processor starting task urgent-1
Urgent processor finished task urgent-1
Normal processor finished task normal-1

收集结果

在前面的示例中，我们依赖控制台输出来验证任务完成情况。但在实际应用中，通常需要以编程方式收集和处理结果。

为了收集这些消息，我们将使用ClosureAgent。我们定义了一个专用主题TASK_RESULTS_TOPIC_TYPE，UrgentProcessor和NormalProcessor都会将结果发布到这个主题。然后ClosureAgent会处理来自该主题的消息。

queue = asyncio.Queue[TaskResponse]()


async def collect_result(_agent: ClosureContext, message: TaskResponse, ctx: MessageContext) -> None:
    await queue.put(message)


runtime.start()

CLOSURE_AGENT_TYPE = "collect_result_agent"
await ClosureAgent.register_closure(
    runtime,
    CLOSURE_AGENT_TYPE,
    collect_result,
    subscriptions=lambda: [TypeSubscription(topic_type=TASK_RESULTS_TOPIC_TYPE, agent_type=CLOSURE_AGENT_TYPE)],
)

await runtime.publish_message(Task(task_id="normal-1"), topic_id=TopicId(type="normal", source="default"))
await runtime.publish_message(Task(task_id="urgent-1"), topic_id=TopicId(type="urgent", source="default"))

await runtime.stop_when_idle()

Normal processor starting task normal-1
Urgent processor starting task urgent-1
Urgent processor finished task urgent-1
Normal processor finished task normal-1

while not queue.empty():
    print(await queue.get())

TaskResponse(task_id='urgent-1', result='Results by Urgent Processor')
TaskResponse(task_id='normal-1', result='Results by Normal Processor')

直接消息

与之前的模式不同，这个模式专注于直接消息。这里我们演示两种发送方式：

• 代理之间的直接消息传递

• 从运行时向特定代理发送消息

需要注意以下几点：

• 消息使用AgentId进行寻址

• 发送方可以期待收到目标代理的响应

• 我们只注册一次WorkerAgent类，但向两个不同的工作线程发送任务

  • 如何实现？如代理生命周期所述，当使用AgentId传递消息时，运行时将获取实例（如果不存在则创建一个）。在本例中，运行时在发送这两条消息时创建了两个工作线程实例。

class WorkerAgent(RoutedAgent):
    @message_handler
    async def on_task(self, message: Task, ctx: MessageContext) -> TaskResponse:
        print(f"{self.id} starting task {message.task_id}")
        await asyncio.sleep(2)  # 模拟工作
        print(f"{self.id} finished task {message.task_id}")
        return TaskResponse(task_id=message.task_id, result=f"Results by {self.id}")


class DelegatorAgent(RoutedAgent):
    def __init__(self, description: str, worker_type: str):
        super().__init__(description)
        self.worker_instances = [AgentId(worker_type, f"{worker_type}-1"), AgentId(worker_type, f"{worker_type}-2")]

    @message_handler
    async def on_task(self, message: Task, ctx: MessageContext) -> TaskResponse:
        print(f"Delegator received task {message.task_id}.")

        subtask1 = Task(task_id="task-part-1")
        subtask2 = Task(task_id="task-part-2")

        worker1_result, worker2_result = await asyncio.gather(
            self.send_message(subtask1, self.worker_instances[0]), self.send_message(subtask2, self.worker_instances[1])
        )

        combined_result = f"Part 1: {worker1_result.result}, " f"Part 2: {worker2_result.result}"
        task_response = TaskResponse(task_id=message.task_id, result=combined_result)
        return task_response

runtime = SingleThreadedAgentRuntime()

await WorkerAgent.register(runtime, "worker", lambda: WorkerAgent("Worker Agent"))
await DelegatorAgent.register(runtime, "delegator", lambda: DelegatorAgent("Delegator Agent", "worker"))

runtime.start()

delegator = AgentId("delegator", "default")
response = await runtime.send_message(Task(task_id="main-task"), recipient=delegator)

print(f"Final result: {response.result}")
await runtime.stop_when_idle()

Delegator received task main-task.
worker/worker-1 starting task task-part-1
worker/worker-2 starting task task-part-2
worker/worker-1 finished task task-part-1
worker/worker-2 finished task task-part-2
Final result: Part 1: Results by worker/worker-1, Part 2: Results by worker/worker-2

额外资源

如果您对并发处理更感兴趣，可以查看智能体混合模式，该模式大量依赖并发智能体。