02 · Providers · OpenAI

What this lesson demonstrates

Everything in module 01 ran on Azure AI Foundry. But an agent.Agent doesn’t care who runs the model — that is the provider’s job. This lesson rebuilds the exact same “Joker” agent, both one-shot and streaming, but through openaiprovider talking to the OpenAI API. Notice how little changes: the constructor and the credentials, and nothing else.

The wiring

The agent is built from an openai.Client and an openaiprovider.AgentConfig that names the model directly:

func newJoker(client openai.Client, model string, mw agent.Middleware) *agent.Agent {
    return openaiprovider.NewAgent(
        client,
        openaiprovider.AgentConfig{
            Model:        model,
            Instructions: "You are good at telling jokes.",
            Config: agent.Config{
                Name:        "Joker",
                Middlewares: []agent.Middleware{mw}, // wraps every run, for logging
            },
        },
    )
}

main runs the prompt twice — first RunText(...).Collect() to get the whole response at once, then the same call with agent.Stream(true) to range over incremental updates. The defaultModel is gpt-4o-mini.

What to notice

• A different credential convention. OpenAI has no azcore.TokenCredential. Instead, openai.NewClient() reads OPENAI_API_KEY (and optional OPENAI_BASE_URL) from the environment — the OpenAI SDK’s own default. main fails fast with a clear message if the key isn’t set, rather than surfacing an opaque 401 mid-run.
• Model lives in AgentConfig. You name the model directly ("gpt-4o-mini") rather than passing a Foundry ModelDeployment(...).
• Middleware is provider-agnostic too. The runLogger — a transparent pass-through that counts runs and prints the prompt, then returns next(...) untouched — is the same middleware from step01_running, reused verbatim.
• Construction is separated from main. newJoker(client, model, mw) lets the offline test build the identical agent with a dummy-keyed client and assert a.Name() == "Joker" and a.ProviderName() == "openai" — with no network, since openai.NewClient() never dials out at construction.

How it maps to the Microsoft Agent Framework

The provider is the whole portability story of the Agent Framework Go SDK. openaiprovider.NewAgent(...) returns the same *agent.Agent that foundryprovider.NewAgent(...) did — so RunText, .Collect(), and streaming are unchanged when you move a workload between OpenAI and Azure AI Foundry. Even swapping NewAgent for NewChatCompletionsAgent (Chat Completions instead of Responses) leaves your call sites and tests untouched.

Run it

OPENAI_API_KEY=sk-... go run ./tutorial/02-agents/providers/openai. The offline test (wiring + middleware) needs no key; TestOpenAI_Live makes the real model call and is gated behind AF_LIVE=1.

Next: step01 · File-Based Skills