agent-framework

name: agent-framework description: 'Build, extend, and debug AI agents in Egroo using the Microsoft Agent Framework (C# .NET). Use for: creating AgentDefinition, adding function tools with AIFunctionFactory, connecting MCP servers, managing agent conversations and session state, wiring new LLM providers (OpenAI, AzureOpenAI, Anthropic, Ollama), adding knowledge items, implementing streaming chat. Also covers AgentRuntimeService internals, BuiltinTools, and McpClientService patterns already in the codebase.' argument-hint: 'Describe what you want to build or fix (e.g. "add a weather tool", "new Ollama agent", "debug streaming")'

Microsoft Agent Framework — Egroo Integration

When to Use

• Creating or modifying an AgentDefinition and its associated CRUD endpoints
• Adding a new function tool (builtin or MCP-backed)
• Wiring a new LLM provider (OpenAI, Azure OpenAI, Anthropic, Ollama)
• Working with agent conversations, session state, or message history
• Implementing or debugging streaming chat (ChatStreamAsync)
• Adding knowledge items injected into the system prompt
• Integrating a new MCP server for tool discovery/invocation

Key Files

For framework details, see ./references/ms-agent-framework.md.

Architecture

Client (Blazor)
  └─ AgentService (HTTP)
       └─ POST /api/v1/Agent/conversations/{id}/send-message
            └─ AgentRuntimeService.ChatAsync / ChatStreamAsync
                 ├─ IAgentRepository  → loads AgentDefinition, knowledge, tools, history
                 ├─ EncryptionService → decrypts agent API key
                 ├─ AIAgent (Microsoft.Agents.AI)
                 │    └─ provider: OpenAI | AzureOpenAI | Anthropic | Ollama
                 ├─ BuiltinTools     → pre-built AIFunctions
                 └─ McpClientService → dynamic MCP tool proxies

Procedure

1. Creating a New Agent (Backend)

1. Model: AgentDefinition (in jihadkhawaja.chat.shared/Models/Agent.cs) holds all config. Required fields: UserId, Name, Provider (LlmProvider enum), Model, ApiKey (stored AES-encrypted), Instructions.

2. Encrypt the API key before saving — EncryptionService.Encrypt(apiKey). Keys are stripped from all API responses.

3. Register via POST /api/v1/Agent — already wired in AgentEndpoint.cs.

4. Repository: All CRUD goes through IAgentRepository → AgentRepository. Register as scoped.

2. Adding a Builtin Function Tool

Add to src/Egroo.Server/Services/BuiltinTools.cs:

public static AIFunction MyNewTool { get; } = AIFunctionFactory.Create(
    ([Description("The input")] string input) =>
    {
        // ... logic ...
        return Task.FromResult("result");
    },
    name: "my_new_tool",
    description: "What this tool does.");

Then include it in AgentRuntimeService alongside existing builtins:

var tools = new List<AIFunction> { BuiltinTools.GetCurrentDatetime, BuiltinTools.MyNewTool };

Seed it for an agent via POST /api/v1/Agent/{agentId}/seed-builtin-tools.

3. Wiring a New LLM Provider

AgentRuntimeService switches on AgentDefinition.Provider. Add a new case:

LlmProvider.Ollama => new OllamaApiClient(new Uri(definition.Endpoint!))
    .AsChatClient(definition.Model)
    .AsAIAgent(instructions: systemPrompt, tools: tools),

Available NuGet packages already in Egroo.Server.csproj:

• Microsoft.Agents.AI.OpenAI — .GetChatClient(model).AsAIAgent(...)
• Microsoft.Agents.AI.Anthropic — AnthropicClient.AsAIAgent(...)
• Microsoft.Extensions.AI.Ollama — OllamaApiClient.AsChatClient(model).AsAIAgent(...)
• Azure.AI.OpenAI — AzureOpenAIClient.GetChatClient(model).AsAIAgent(...)

4. Adding Knowledge Items

Knowledge is injected into the system prompt by AgentRuntimeService:

var knowledgeItems = await _agentRepo.GetAgentKnowledgeAsync(agentId);
var systemPrompt = definition.Instructions + "\n\n" +
    string.Join("\n", knowledgeItems.Where(k => k.IsEnabled).Select(k => k.Content));

Add via POST /api/v1/Agent/{agentId}/knowledge.

5. Connecting an MCP Server

1. Register: POST /api/v1/Agent/{agentId}/mcp-servers with endpoint + optional API key.
2. Discover tools: POST /api/v1/Agent/{agentId}/mcp-servers/discover-tools — calls McpClientService.DiscoverToolsAsync and persists to AgentTool table with Source = Mcp.
3. At runtime, AgentRuntimeService wraps each MCP tool as an AIFunction that proxies through McpClientService.CallToolAsync.

McpClientService uses standard JSON-RPC 2.0 (tools/list, tools/call). API keys are sent as Authorization: Bearer headers.

6. Running a Conversation Turn

// Non-streaming
var response = await _agentRuntimeService.ChatAsync(userId, conversationId, "Hello");

// Streaming (IAsyncEnumerable<string>)
await foreach (var token in _agentRuntimeService.ChatStreamAsync(userId, conversationId, "Hello"))
{
    // push token to client via SignalR or SSE
}

Session state and message history are persisted per AgentConversation.Id via IAgentRepository.

7. Adding a New API Endpoint

Follow the Minimal API group pattern in AgentEndpoint.cs:

group.MapPost("/{agentId}/my-feature", async (IAgentRepository repo, Guid agentId, MyRequest req, HttpContext ctx) =>
{
    // use repo, return Results.Ok/NotFound
})
.RequireAuthorization();

All agent routes require JWT auth (RequireAuthorization()) and inherit the "Api" rate limit policy.

Security Checklist

• API keys are always encrypted with EncryptionService before DB storage
• API keys are stripped from all responses (check AgentEndpoint.cs response mapping)
• MCP server endpoints should be validated/allowlisted before calling in production
• GetConnectorUserId() from BaseRepository is used to scope agents to the authenticated user
• Never log decrypted API keys

Common Pitfalls

• LlmProvider.Ollama requires an Endpoint field on AgentDefinition — validate it non-null before calling.
• Conversation SessionState is stored as serialized JSON in the DB — full history is loaded on every turn. For long conversations, consider truncation.
• MCP tool discover is manual — tools are not auto-refreshed; call the discover endpoint after updating an MCP server.
• jihadkhawaja.chat.shared changes require rebuilding shared → server/client in order before tests pass.
• Streaming endpoint must be called with HttpCompletionOption.ResponseHeadersRead from the client.