Azure · OpenAI Chat Completions

What this lesson demonstrates

This lesson backs the agent with Azure OpenAI’s Chat Completions API. It uses the OpenAI Go SDK client pointed at an Azure OpenAI resource, wrapped by openaiprovider.NewChatCompletionsAgent. Where the foundry_model lesson used the generic NewAgent (Responses API), this one calls the chat-completions constructor explicitly — the same provider, a different endpoint shape underneath. The agent surface — RunText, Collect, Stream — is identical; only the construction and the deployment convention differ.

The lesson also runs the agent twice: once one-shot to completion, once with agent.Stream(true) to show incremental updates ranged over as a Go 1.23+ range-over-func iterator.

One real excerpt

func newJoker(endpoint, model string, cred azcore.TokenCredential) *agent.Agent {
    return openaiprovider.NewChatCompletionsAgent(
        openai.NewClient(
            option.WithBaseURL(endpoint),    // point the OpenAI client at Azure
            azure.WithTokenCredential(cred), // authenticate with an Azure token
        ),
        openaiprovider.AgentConfig{
            Model:        model, // on Azure this is the deployment name
            Instructions: "You are good at telling jokes.",
            Config: agent.Config{
                Name:        "Joker",
                Middlewares: []agent.Middleware{&runLogger{name: "Joker"}},
            },
        },
    )
}

What to notice — the gotcha

On Azure, the Model field is your deployment name, not a model ID. When you deploy a model in the Azure portal you give the deployment a name, and that name — not gpt-4o-mini as a catalog identifier — is what you call. The lesson reads AZURE_OPENAI_DEPLOYMENT_NAME, falling back to gpt-4o-mini as the default deployment name; the endpoint comes from AZURE_OPENAI_ENDPOINT (e.g. https://<resource>.openai.azure.com).

Note this constructor uses azure.WithTokenCredential(cred) with the default scope — it does not pass the ai.azure.com scope that the foundry_model lesson needed. That’s the difference between an Azure OpenAI resource and a Foundry model resource. The credential (DefaultAzureCredential, i.e. your az login session) is lazy, so the offline test builds the agent with a fake credential and asserts the middleware is wired without any network.

How it maps to the Agent Framework

openaiprovider offers two constructors — NewChatCompletionsAgent and NewResponsesAgent — because Azure OpenAI exposes both the older Chat Completions API and the newer Responses API, and they have different request shapes. Choosing the right constructor is the entire provider decision; the agent.Config (name, instructions, middleware) is shared. This lets one codebase target Chat Completions for a legacy deployment and Responses for a newer one, with only the constructor line changing.

Run it

go run ./tutorial/02-agents/providers/azure/openai_chat_completion

Live runs need az login plus AZURE_OPENAI_ENDPOINT (and optional AZURE_OPENAI_DEPLOYMENT_NAME). The structural test runs offline with a fake credential; live calls are gated behind AF_LIVE=1.

Next: 02 · Providers · Azure · OpenAI Responses