Semantic Kernel - Multi agent systems

Yesterday I talked about the new agent abstraction in Semantic Kernel and how it can simplify the steps required to build your own AI agent.  But what could be better than having one agent? Multiple agents of course!

And that is exactly what was recently introduced as a preview in Semantic Kernel.

As explained in this blog post, there are multiple ways that multiple agents can work together. The simplest way is as a group chat where multiple agents can talk back-and-forth with each other. To avoid that these agents get stuck in a loop this is combined with a custom termination strategy that specifies when the conversation is over.

Here is a small example.

I start with the default Semantic Kernel configuration to create a kernel instance:

	HttpClient client = new HttpClient();
	client.Timeout = TimeSpan.FromMinutes(2);
	var builder = Kernel.CreateBuilder()
	.AddOpenAIChatCompletion(
	modelId: "phi3.5:latest", apiKey: null, endpoint: new Uri("http://localhost:11434"), httpClient: client);
	builder.Services.AddLogging(c => c.SetMinimumLevel(LogLevel.Trace).AddDebug());
	// Build the kernel
	Kernel kernel = builder.Build();

view raw Step1.cs hosted with ❤ by GitHub

Now I define the instructions for the different agents and create them:

	// Define the instructions for the different agents
	string Editor = """
	You are an editor which will take a text and make it easier to understand. Ensure the key information is preserved while using clear, concise language. Remove unnecessary jargon and complex sentence structures, but maintain the original meaning and tone.
	""";
	string SpellingCorrector = """
	You are a spelling correcot. You review a text and correct any spelling mistakes. Ensure all words are spelled correctly without altering the meaning or structure of the original text.
	""";
	string ChiefEditor = """
	You are a chief editor which will review a text before it can be printed.
	If the text is OK, just respond "approve".
	""";
	#pragma warning disable SKEXP0110, SKEXP0001 // Rethrow to preserve stack details
	ChatCompletionAgent EditorAgent =
	new()
	{
	Instructions = Editor,
	Name = "EditorAgent",
	Kernel = kernel
	};
	ChatCompletionAgent SpellingAgent =
	new()
	{
	Instructions = SpellingCorrector,
	Name = "SpellingAgent",
	Kernel = kernel
	};
	ChatCompletionAgent ChiefEditorAgent =
	new()
	{
	Instructions = ChiefEditor,
	Name = "ChiefEditorAgent",
	Kernel = kernel
	};

view raw Step2.cs hosted with ❤ by GitHub

Remark: Notice that I can use different kernels with different models if I want to.

To make sure that the conversation is ended I need to specify a TerminationStrategy:

	sealed class ApprovalTerminationStrategy : TerminationStrategy
	{
	// Terminate when the final message contains the term "approve"
	protected override Task<bool> ShouldAgentTerminateAsync(Agent agent, IReadOnlyList<ChatMessageContent> history, CancellationToken cancellationToken)
	=> Task.FromResult(history[history.Count - 1].Content?.Contains("approve", StringComparison.OrdinalIgnoreCase) ?? false);
	}

view raw Step3.cs hosted with ❤ by GitHub

As a last step I need to bring the multiple agents together in a group chat:

	AgentGroupChat groupChat =
	new(EditorAgent, SpellingAgent, ChiefEditorAgent)
	{
	ExecutionSettings =
	new()
	{
	TerminationStrategy =
	new ApprovalTerminationStrategy()
	{
	Agents = [ChiefEditorAgent],
	MaximumIterations = 3,
	}
	}
	};

view raw Step4.cs hosted with ❤ by GitHub

Now I can start the conversation by providing some input:

	string input = """
	Can you help me edit the following text: I like to write very complex sentences with lots of jargan and big words. I hope you can help me make it easier to understand.
	""";
	groupChat.AddChatMessage(new ChatMessageContent(AuthorRole.User, input));
	Console.WriteLine($"# {AuthorRole.User}: '{input}'");
	await foreach (var content in groupChat.InvokeAsync())
	{
	Console.WriteLine($"# {content.Role} - {content.AuthorName ?? "*"}: '{content.Content}'");
	}

view raw Step5.cs hosted with ❤ by GitHub

That’s it!

The full example:

	HttpClient client = new HttpClient();
	client.Timeout = TimeSpan.FromMinutes(2);
	var builder = Kernel.CreateBuilder()
	.AddOpenAIChatCompletion(
	modelId: "phi3.5:latest", apiKey: null, endpoint: new Uri("http://localhost:11434"), httpClient: client);
	builder.Services.AddLogging(c => c.SetMinimumLevel(LogLevel.Trace).AddDebug());
	// Build the kernel
	Kernel kernel = builder.Build();
	// Define the instructions for the different agents
	string Editor = """
	You are an editor which will take a text and make it easier to understand. Ensure the key information is preserved while using clear, concise language. Remove unnecessary jargon and complex sentence structures, but maintain the original meaning and tone.
	""";
	string SpellingCorrector = """
	You are a spelling correcot. You review a text and correct any spelling mistakes. Ensure all words are spelled correctly without altering the meaning or structure of the original text.
	""";
	string ChiefEditor = """
	You are a chief editor which will review a text before it can be printed.
	If the text is OK, just respond "approve".
	""";
	#pragma warning disable SKEXP0110, SKEXP0001 // Rethrow to preserve stack details
	ChatCompletionAgent EditorAgent =
	new()
	{
	Instructions = Editor,
	Name = "EditorAgent",
	Kernel = kernel
	};
	ChatCompletionAgent SpellingAgent =
	new()
	{
	Instructions = SpellingCorrector,
	Name = "SpellingAgent",
	Kernel = kernel
	};
	ChatCompletionAgent ChiefEditorAgent =
	new()
	{
	Instructions = ChiefEditor,
	Name = "ChiefEditorAgent",
	Kernel = kernel
	};
	AgentGroupChat groupChat =
	new(EditorAgent, SpellingAgent, ChiefEditorAgent)
	{
	ExecutionSettings =
	new()
	{
	TerminationStrategy =
	new ApprovalTerminationStrategy()
	{
	Agents = [ChiefEditorAgent],
	MaximumIterations = 3,
	}
	}
	};
	string input = """
	Can you help me edit the following text: I like to write very complex sentences with lots of jargan and big words. I hope you can help me make it easier to understand.
	""";
	groupChat.AddChatMessage(new ChatMessageContent(AuthorRole.User, input));
	Console.WriteLine($"# {AuthorRole.User}: '{input}'");
	await foreach (var content in groupChat.InvokeAsync())
	{
	Console.WriteLine($"# {content.Role} - {content.AuthorName ?? "*"}: '{content.Content}'");
	}
	sealed class ApprovalTerminationStrategy : TerminationStrategy
	{
	// Terminate when the final message contains the term "approve"
	protected override Task<bool> ShouldAgentTerminateAsync(Agent agent, IReadOnlyList<ChatMessageContent> history, CancellationToken cancellationToken)
	=> Task.FromResult(history[history.Count - 1].Content?.Contains("approve", StringComparison.OrdinalIgnoreCase) ?? false);
	}

view raw MultiAgent.cs hosted with ❤ by GitHub

More information

Exploring Multi-Agent AI Systems (microsoft.com)

Introducing enterprise multi-agent support in Semantic Kernel | Semantic Kernel (microsoft.com)