02 · Providers · Azure · OpenAI Responses

What this lesson demonstrates

This lesson swaps the provider onto Azure OpenAI’s Responses API. The agent contract is identical to the very first hello_agent lesson — instructions plus a model, RunText, Collect — but the constructor is openaiprovider.NewResponsesAgent and the client is an openai.Client pointed at your Azure OpenAI endpoint with an Azure token credential.

The interesting twist is that it builds the agent twice to show one knob: DisableStoreOutput. By default the Responses service stores conversation state server-side; set DisableStoreOutput = true (the API’s "store": false) and Agent Framework keeps the chat history locally instead.

One real excerpt

func newJoker(client openai.Client, model string, disableStore bool, mw agent.Middleware) *agent.Agent {
    return openaiprovider.NewResponsesAgent(
        client,
        openaiprovider.AgentConfig{
            Model:              model,
            Instructions:       "You are good at telling jokes.",
            DisableStoreOutput: disableStore,
            Config: agent.Config{
                Name:        "Joker",
                Middlewares: []agent.Middleware{mw},
            },
        },
    )
}

main builds one agent with disableStore=false (server-side state) and a second with true (local history).

What to notice — the gotcha

DisableStoreOutput decides who owns the conversation history. With the default (false), the Responses service persists conversation state on the server, and follow-up turns reference it by ID. Flip it to true and Agent Framework manages the history in-process instead — useful when you don’t want the service retaining conversation state, or when you’re stitching history together yourself. The two agents in the lesson are otherwise identical; only this one boolean differs.

The endpoint comes from AZURE_OPENAI_ENDPOINT and the deployment from AZURE_OPENAI_DEPLOYMENT_NAME (default gpt-4o-mini — remember that on Azure this is the deployment name). newClient builds the openai.Client with option.WithBaseURL and azure.WithTokenCredential(cred); the credential is DefaultAzureCredential, lazy, so the offline test builds both agent variants with a fake credential and asserts their wiring with no network.

How it maps to the Agent Framework

The Responses API is Azure OpenAI’s stateful conversation surface, and DisableStoreOutput is the framework’s clean handle on the server-side-vs-local-state tradeoff — a decision that would otherwise mean reaching into raw request parameters. Pairing this with the openai_chat_completion lesson gives you the full Azure OpenAI picture: NewChatCompletionsAgent for the stateless Chat Completions API, NewResponsesAgent for the stateful Responses API, both from the same openaiprovider, both exposing the identical agent.Agent interface.

Run it

AZURE_OPENAI_ENDPOINT=https://<res>.openai.azure.com go run ./tutorial/02-agents/providers/azure/openai_responses

Live runs need AZURE_OPENAI_ENDPOINT and an Azure login (az login). The structural test runs offline with a fake credential and asserts both variants; live calls are gated behind AF_LIVE=1.

Next: step01 · Basic Foundry Provider