A2A Protocol

Overview

A2A (Agent-to-Agent) is an open protocol by Google that enables AI agents built on different frameworks to discover each other, negotiate capabilities, and collaborate on tasks. The .NET SDK (SharpA2A) provides a strongly-typed client and server implementation with support for agent cards, task lifecycle management, streaming via SSE, and push notifications.

NuGet Packages

bash

dotnet add package SharpA2A.Core
dotnet add package SharpA2A.AspNetCore

Agent Cards

An agent card is a JSON metadata document that advertises an agent's capabilities, skills, and endpoint. Agents discover each other by fetching /.well-known/agent.json.

csharp

using SharpA2A.Core;

var agentCard = new AgentCard
{
    Name = "ResearchAgent",
    Description = "Searches academic papers and summarizes findings",
    Url = "https://research-agent.example.com",
    Version = "1.0.0",
    Capabilities = new AgentCapabilities
    {
        Streaming = true,
        PushNotifications = false
    },
    Skills = new List<AgentSkill>
    {
        new AgentSkill
        {
            Id = "paper-search",
            Name = "Paper Search",
            Description = "Search and summarize academic papers on a given topic",
            Tags = new List<string> { "research", "academic", "papers" },
            Examples = new List<string>
            {
                "Find recent papers on transformer architectures",
                "Summarize the key findings in federated learning research"
            }
        }
    },
    DefaultInputModes = new List<string> { "text/plain" },
    DefaultOutputModes = new List<string> { "text/plain", "text/markdown" }
};

A2A Server with ASP.NET Core

csharp

using SharpA2A.Core;
using SharpA2A.AspNetCore;

var builder = WebApplication.CreateBuilder(args);

var agentCard = new AgentCard
{
    Name = "SummaryAgent",
    Description = "Summarizes long documents into concise bullet points",
    Url = "https://localhost:5001",
    Version = "1.0.0",
    Capabilities = new AgentCapabilities { Streaming = true }
};

builder.Services.AddA2AServer(agentCard, async (taskRequest, cancellationToken) =>
{
    var userMessage = taskRequest.Messages.LastOrDefault()?.Parts
        .OfType<TextPart>().FirstOrDefault()?.Text ?? "";

    var summary = await SummarizeDocumentAsync(userMessage, cancellationToken);

    return new TaskResponse
    {
        Status = new TaskStatus { State = TaskState.Completed },
        Artifacts = new List<Artifact>
        {
            new Artifact
            {
                Name = "summary",
                Parts = new List<Part> { new TextPart { Text = summary } }
            }
        }
    };
});

var app = builder.Build();
app.MapA2AEndpoints();
app.Run();

A2A Client

csharp

using SharpA2A.Core;

var client = new A2AClient("https://research-agent.example.com");

// Discover agent capabilities
var card = await client.GetAgentCardAsync();
Console.WriteLine($"Agent: {card.Name}, Skills: {card.Skills.Count}");

// Send a task
var taskRequest = new TaskSendParams
{
    Id = Guid.NewGuid().ToString(),
    Message = new Message
    {
        Role = "user",
        Parts = new List<Part>
        {
            new TextPart { Text = "Summarize recent advances in quantum computing" }
        }
    }
};

var response = await client.SendTaskAsync(taskRequest);
Console.WriteLine($"Task state: {response.Status.State}");

foreach (var artifact in response.Artifacts ?? Enumerable.Empty<Artifact>())
{
    foreach (var part in artifact.Parts.OfType<TextPart>())
    {
        Console.WriteLine(part.Text);
    }
}

Streaming Responses

csharp

var streamRequest = new TaskSendParams
{
    Id = Guid.NewGuid().ToString(),
    Message = new Message
    {
        Role = "user",
        Parts = new List<Part>
        {
            new TextPart { Text = "Write a detailed analysis of climate data trends" }
        }
    }
};

await foreach (var update in client.SendTaskStreamingAsync(streamRequest))
{
    if (update is TaskStatusUpdateEvent statusUpdate)
    {
        Console.WriteLine($"Status: {statusUpdate.Status.State}");
        if (statusUpdate.Status.Message is { } msg)
        {
            foreach (var part in msg.Parts.OfType<TextPart>())
            {
                Console.Write(part.Text);
            }
        }
    }
    else if (update is TaskArtifactUpdateEvent artifactUpdate)
    {
        foreach (var part in artifactUpdate.Artifact.Parts.OfType<TextPart>())
        {
            Console.Write(part.Text);
        }
    }
}

Task Lifecycle

The A2A protocol defines a clear task state machine:

State	Description
`Submitted`	Task received by the agent, not yet started
`Working`	Agent is actively processing the task
`InputRequired`	Agent needs additional input from the caller
`Completed`	Task finished successfully with artifacts
`Failed`	Task encountered an unrecoverable error
`Canceled`	Task was canceled by the caller or agent

csharp

// Handling multi-turn conversations (InputRequired state)
var response = await client.SendTaskAsync(initialRequest);

while (response.Status.State == TaskState.InputRequired)
{
    var followUpMessage = response.Status.Message?.Parts
        .OfType<TextPart>().FirstOrDefault()?.Text;
    Console.WriteLine($"Agent asks: {followUpMessage}");

    var userInput = Console.ReadLine()!;
    var followUp = new TaskSendParams
    {
        Id = response.Id,
        Message = new Message
        {
            Role = "user",
            Parts = new List<Part> { new TextPart { Text = userInput } }
        }
    };
    response = await client.SendTaskAsync(followUp);
}

Multi-Agent Orchestration

csharp

public class OrchestratorAgent
{
    private readonly Dictionary<string, A2AClient> _agents = new();

    public async Task RegisterAgentAsync(string agentUrl)
    {
        var client = new A2AClient(agentUrl);
        var card = await client.GetAgentCardAsync();
        _agents[card.Name] = client;
    }

    public async Task<string> DelegateTaskAsync(string skillTag, string prompt)
    {
        // Find an agent that has the required skill
        foreach (var (name, client) in _agents)
        {
            var card = await client.GetAgentCardAsync();
            var matchingSkill = card.Skills
                .FirstOrDefault(s => s.Tags.Contains(skillTag));

            if (matchingSkill is not null)
            {
                var request = new TaskSendParams
                {
                    Id = Guid.NewGuid().ToString(),
                    Message = new Message
                    {
                        Role = "user",
                        Parts = new List<Part> { new TextPart { Text = prompt } }
                    }
                };

                var result = await client.SendTaskAsync(request);
                return result.Artifacts?.FirstOrDefault()?.Parts
                    .OfType<TextPart>().FirstOrDefault()?.Text
                    ?? "No result";
            }
        }
        throw new InvalidOperationException($"No agent found with skill tag '{skillTag}'");
    }
}

Best Practices

•Publish an agent card at /.well-known/agent.json with accurate skill descriptions and supported input/output modes so callers can discover and filter agents programmatically.
•Use unique, deterministic task IDs (e.g., combining a request correlation ID with a timestamp) to enable idempotent retries and task tracking across agents.
•Implement the InputRequired state for multi-turn interactions instead of trying to collect all information upfront in a single prompt.
•Return structured Artifact objects with explicit MIME types rather than embedding results in status messages, so downstream agents can parse outputs reliably.
•Enable streaming (Capabilities.Streaming = true) for long-running tasks so callers receive incremental progress rather than blocking on a single response.
•Version your agent cards and skill schemas; include a Version field and avoid breaking changes to skill IDs or parameter formats across releases.
•Validate incoming task messages against expected input modes before processing and return a Failed state with a descriptive error when validation fails.
•Register agent clients via dependency injection with AddA2AServer rather than constructing clients manually, to enable proper lifetime management and testability.
•Implement cancellation token propagation through the entire task handler chain so that canceled tasks release resources promptly.
•Log task state transitions with correlation IDs to enable distributed tracing across multi-agent orchestrations.