Why should you care?

Testability and swapping OpenAI for Azure OpenAI per environment depend on constructor injection, not static SDK calls.

See it live — copy this example

Paste into an ASP.NET Core 8+ / AgentNest project, then run with dotnet run (set your API keys in user-secrets).

// AgentNest.Api/DependencyInjection/AiServiceCollectionExtensions.cs
public static class AiServiceCollectionExtensions
{
    public static IServiceCollection AddAgentNestAi(this IServiceCollection services, IConfiguration config)
    {
        services.AddSingleton<IChatClient>(sp =>
            sp.GetRequiredService<IChatClient>()); // registered by AddOpenAIChatClient
        services.AddScoped<ICrmCopilotAgent, CrmCopilotAgent>();
        services.AddScoped<IHospitalTriageAgent, HospitalTriageAgent>();
        return services;
    }
}

// Program.cs: builder.Services.AddAgentNestAi(builder.Configuration);

What happened?

• AddAgentNestAi groups AI registrations.
• Scoped agents get per-request isolation; chat clients stay singletons wrapping HTTP connections.

Practice next

1. Create AiServiceCollectionExtensions in AgentNest.Api.
2. Register agents as scoped; avoid transient Kernel builds.
3. Replace IChatClient with a fake in unit tests via WebApplicationFactory.
4. Add keyed services for OpenAI vs Azure per tenant feature flag.
5. Register IEmbeddingGenerator alongside IChatClient.

Remember

Group AI DI in extension methods for AgentNest modules. Match lifetimes: singleton clients, scoped agents. Inject interfaces so tests swap fakes easily.

Swap model provider in CI

Pipeline tests must not call real OpenAI on every PR.

Outcome: DI replaces IChatClient with TestChatClient in WebApplicationFactory fixtures.