Orchestration magentique

L’orchestration Magentic est conçue sur la base du système Magentic-One inventé par AutoGen. Il s’agit d’un modèle multi-agent à usage général flexible conçu pour les tâches complexes et ouvertes qui nécessitent une collaboration dynamique. Dans ce modèle, un responsable Magentic dédié coordonne une équipe d’agents spécialisés, en sélectionnant l’agent qui doit agir ensuite en fonction du contexte en évolution, de la progression des tâches et des fonctionnalités de l’agent.

Le gestionnaire Magentic gère un contexte partagé, suit la progression et adapte le flux de travail en temps réel. Cela permet au système de décomposer les problèmes complexes, de déléguer des tâches subordonnées et d’affiner de manière itérative des solutions via la collaboration de l’agent. L’orchestration est particulièrement adaptée aux scénarios où le chemin de la solution n’est pas connu à l’avance et peut nécessiter plusieurs séries de raisonnement, de recherche et de calcul.

Pour en savoir plus sur le modèle, par exemple quand utiliser le modèle ou quand éviter le modèle dans votre charge de travail, consultez orchestration Magentic.

Cas d’usage courants

Un utilisateur demande un rapport complet comparant l’efficacité énergétique et les émissions de CO² de différents modèles Machine Learning. Le responsable Magentic affecte d’abord un agent de recherche pour collecter des données pertinentes, puis délègue l’analyse et le calcul à un agent coder. Le responsable coordonne plusieurs séries de recherches et de calculs, agrège les résultats et produit un rapport détaillé et structuré en tant que sortie finale.

Ce que vous allez apprendre

• Guide pratique pour définir et configurer des agents pour l’orchestration Magentic
• Comment configurer un responsable Magentic pour coordonner la collaboration de l’agent
• Fonctionnement du processus d’orchestration, y compris la planification, le suivi des progrès et la synthèse finale des réponses

using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Agents;
using Microsoft.SemanticKernel.Agents.AzureAI;
using Microsoft.SemanticKernel.Agents.Magentic;
using Microsoft.SemanticKernel.Agents.Orchestration;
using Microsoft.SemanticKernel.Agents.Runtime.InProcess;
using Microsoft.SemanticKernel.ChatCompletion;
using Microsoft.SemanticKernel.Connectors.OpenAI;
using Azure.AI.Agents.Persistent;
using Azure.Identity;

// Helper function to create a kernel with chat completion
public static Kernel CreateKernelWithChatCompletion(...)
{
    ...
}

// Create a kernel with OpenAI chat completion for the research agent
Kernel researchKernel = CreateKernelWithChatCompletion("gpt-4o-search-preview");
ChatCompletionAgent researchAgent = new ChatCompletionAgent {
    Name = "ResearchAgent",
    Description = "A helpful assistant with access to web search. Ask it to perform web searches.",
    Instructions = "You are a Researcher. You find information without additional computation or quantitative analysis.",
    Kernel = researchKernel,
};

// Create a persistent Azure AI agent for code execution
PersistentAgentsClient agentsClient = AzureAIAgent.CreateAgentsClient(endpoint, new AzureCliCredential());
PersistentAgent definition = await agentsClient.Administration.CreateAgentAsync(
    modelId,
    name: "CoderAgent",
    description: "Write and executes code to process and analyze data.",
    instructions: "You solve questions using code. Please provide detailed analysis and computation process.",
    tools: [new CodeInterpreterToolDefinition()]);
AzureAIAgent coderAgent = new AzureAIAgent(definition, agentsClient);

Kernel managerKernel = CreateKernelWithChatCompletion("o3-mini");
StandardMagenticManager manager = new StandardMagenticManager(
    managerKernel.GetRequiredService<IChatCompletionService>(),
    new OpenAIPromptExecutionSettings())
{
    MaximumInvocationCount = 5,
};

ChatHistory history = [];

ValueTask responseCallback(ChatMessageContent response)
{
    history.Add(response);
    return ValueTask.CompletedTask;
}

MagenticOrchestration orchestration = new MagenticOrchestration(
    manager,
    researchAgent,
    coderAgent)
{
    ResponseCallback = responseCallback,
};

InProcessRuntime runtime = new InProcessRuntime();
await runtime.StartAsync();

string input = @"I am preparing a report on the energy efficiency of different machine learning model architectures.\nCompare the estimated training and inference energy consumption of ResNet-50, BERT-base, and GPT-2 on standard datasets (e.g., ImageNet for ResNet, GLUE for BERT, WebText for GPT-2). Then, estimate the CO2 emissions associated with each, assuming training on an Azure Standard_NC6s_v3 VM for 24 hours. Provide tables for clarity, and recommend the most energy-efficient model per task type (image classification, text classification, and text generation).";
var result = await orchestration.InvokeAsync(input, runtime);

string output = await result.GetValueAsync(TimeSpan.FromSeconds(300));
Console.WriteLine($"\n# RESULT: {output}");
Console.WriteLine("\n\nORCHESTRATION HISTORY");
foreach (ChatMessageContent message in history)
{
    // Print each message
    Console.WriteLine($"# {message.Role} - {message.AuthorName}: {message.Content}");
}

await runtime.RunUntilIdleAsync();

# RESULT: ```markdown
# Report: Energy Efficiency of Machine Learning Model Architectures

This report assesses the energy consumption and related CO₂ emissions for three popular ...

ORCHESTRATION HISTORY

# Assistant - ResearchAgent: Comparing the energy efficiency of different machine learning ...

# assistant - CoderAgent: Below are tables summarizing the approximate energy consumption and ...

# assistant - CoderAgent: The estimates provided in our tables align with a general understanding ...

# assistant - CoderAgent: Here's the updated structure for the report integrating both the ...

from semantic_kernel.agents import ChatCompletionAgent, OpenAIAssistantAgent
from semantic_kernel.connectors.ai.open_ai import OpenAIChatCompletion

research_agent = ChatCompletionAgent(
    name="ResearchAgent",
    description="A helpful assistant with access to web search. Ask it to perform web searches.",
    instructions="You are a Researcher. You find information without additional computation or quantitative analysis.",
    service=OpenAIChatCompletion(ai_model_id="gpt-4o-search-preview"),
)

# Create an OpenAI Assistant agent with code interpreter capability
client, model = OpenAIAssistantAgent.setup_resources()
code_interpreter_tool, code_interpreter_tool_resources = OpenAIAssistantAgent.configure_code_interpreter_tool()
definition = await client.beta.assistants.create(
    model=model,
    name="CoderAgent",
    description="A helpful assistant that writes and executes code to process and analyze data.",
    instructions="You solve questions using code. Please provide detailed analysis and computation process.",
    tools=code_interpreter_tool,
    tool_resources=code_interpreter_tool_resources,
)
coder_agent = OpenAIAssistantAgent(
    client=client,
    definition=definition,
)

from semantic_kernel.agents import StandardMagenticManager
from semantic_kernel.connectors.ai.open_ai import OpenAIChatCompletion

manager = StandardMagenticManager(chat_completion_service=OpenAIChatCompletion())

from semantic_kernel.contents import ChatMessageContent

def agent_response_callback(message: ChatMessageContent) -> None:
    print(f"**{message.name}**\n{message.content}")

from semantic_kernel.agents import MagenticOrchestration

magentic_orchestration = MagenticOrchestration(
    members=[research_agent, coder_agent],
    manager=manager,
    agent_response_callback=agent_response_callback,
)

from semantic_kernel.agents.runtime import InProcessRuntime

runtime = InProcessRuntime()
runtime.start()

orchestration_result = await magentic_orchestration.invoke(
    task=(
        "I am preparing a report on the energy efficiency of different machine learning model architectures. "
        "Compare the estimated training and inference energy consumption of ResNet-50, BERT-base, and GPT-2 "
        "on standard datasets (e.g., ImageNet for ResNet, GLUE for BERT, WebText for GPT-2). "
        "Then, estimate the CO2 emissions associated with each, assuming training on an Azure Standard_NC6s_v3 VM "
        "for 24 hours. Provide tables for clarity, and recommend the most energy-efficient model "
        "per task type (image classification, text classification, and text generation)."
    ),
    runtime=runtime,
)

value = await orchestration_result.get()
print(f"\nFinal result:\n{value}")

await runtime.stop_when_idle()

**ResearchAgent**
Estimating the energy consumption and associated CO₂ emissions for training and inference of ResNet-50, BERT-base...

**CoderAgent**
Here is the comparison of energy consumption and CO₂ emissions for each model (ResNet-50, BERT-base, and GPT-2)
over a 24-hour period:

| Model     | Training Energy (kWh) | Inference Energy (kWh) | Total Energy (kWh) | CO₂ Emissions (kg) |
|-----------|------------------------|------------------------|---------------------|---------------------|
| ResNet-50 | 21.11                  | 0.08232                | 21.19232            | 19.50               |
| BERT-base | 0.048                  | 0.23736                | 0.28536             | 0.26                |
| GPT-2     | 42.22                  | 0.35604                | 42.57604            | 39.17               |

...

Final result:
Here is the comprehensive report on energy efficiency and CO₂ emissions for ResNet-50, BERT-base, and GPT-2 models...

Étapes suivantes