OpenAI Agents SDK (JS)

@openai/agents. It uses the Responses API.

Connect

Build a model object, so the SDK sends the id as given (its string lookup would read anthropic/... as a provider prefix). Turn tracing off with setTracingDisabled(true), or the SDK posts traces to OpenAI:

/** The SDK model for one id. A model object sends the id as given: the SDK's own string lookup reads `anthropic/...` as a provider prefix. */
export function modelFor(client: OpenAI, model: string): OpenAIResponsesModel {
  return new OpenAIResponsesModel(client, model)
}

Chat

const agent = new Agent({ name: "pong", model: modelFor(client, model), instructions: "Answer tersely." })
const result = await run(agent, "Reply with exactly: pong")

Tool calls

The SDK runs the tool and makes the follow-up request itself.

const getWeather = tool({
  name: "get_weather",
  description: "Current weather for a city",
  parameters: z.object({ city: z.string().describe("City name") }),
  execute: async ({ city }) => {
    cities.push(city)
    return { temp_c: 18, sky: "cloudy" }
  },
})
const agent = new Agent({ name: "weather", model: modelFor(client, model), instructions: "Use the tool to answer.", tools: [getWeather] })
const result = await run(agent, PROMPT)

Streaming

run(agent, input, { stream: true }) yields raw_model_stream_event items; text arrives as output_text_delta.

const agent = new Agent({ name: "pong", model: modelFor(client, model), instructions: "Answer tersely." })
const stream = await run(agent, "Reply with exactly: pong", { stream: true })
let text = ""
let status: unknown
for await (const event of stream) {
  if (event.type === "raw_model_stream_event" && event.data.type === "output_text_delta") text += event.data.delta
  // The streamed response's status comes on its last event, not on the rawResponses entry.
  if (event.type === "raw_model_stream_event" && event.data.type === "response_done") status = event.data.response.providerData?.status
}

Images

// Any data:image/png;base64 URL works. This is a 32px red PNG (examples/_shared/red32.dataurl.txt).
const red = readFileSync(new URL("../_shared/red32.dataurl.txt", import.meta.url), "utf8").trim()
const agent = new Agent({ name: "color", model: modelFor(client, model), instructions: "Answer in one word." })
const result = await run(agent, [
  {
    role: "user",
    content: [
      { type: "input_text", text: "What color is this image? One word." },
      { type: "input_image", image: red },
    ],
  },
])

Structured output

outputType takes a zod schema:

const Meal = z.object({ calories: z.number(), protein_g: z.number() })
const agent = new Agent({ name: "meal", model: modelFor(client, model), instructions: "Estimate the meal.", outputType: Meal })
const result = await run(agent, "Estimate: 2 boiled eggs and a slice of toast.")

Limits

• A wrong key throws AuthenticationError (401) and an unknown model NotFoundError (404).
• The SDK doesn't expose the wire finish_reason, and has no model-list call.