Build an agent in 5 minutes

Agents bundle a model, instructions, and optional tools. They expose the two AI SDK execution modes: generate for single responses and stream for token streams.

1. Instantiate an agent

import { Agent, scaleway } from "@ai_kit/core";

const assistant = new Agent({
  name: "docs-assistant",
  instructions: "You help developers understand the AI Kit platform.",
  model: scaleway("gpt-oss-120b"),
});

name identifies the agent for logging and telemetry.
instructions provide the default system prompt (override per call with system).
model expects a LanguageModel from the AI SDK. Pick the provider you need.

2. Generate a single response

const result = await assistant.generate({
  prompt: "Explain the difference between generate and stream in AI Kit.",
});

console.log(result.text);

The return value mirrors ai.generateText:

text holds the final answer.
response exposes raw metadata (messages, usage, tool calls, …).
loopTool tells you whether the tool loop was triggered.

Conversation mode

await assistant.generate({
  messages: [
    { role: "user", content: "Give me three tutorial ideas." },
  ],
  maxOutputTokens: 256,
});

When you pass messages, AI Kit automatically injects the agent instructions as a system message.

3. Enable telemetry

const supportAgent = new Agent({
  name: "support-assistant",
  model: scaleway("gpt-oss-120b"),
  telemetry: true,
});

The telemetry flag enables Langfuse export when instrumentation is configured. Call agent.withTelemetry(false) to disable it temporarily.

4. Keep exploring

Add tools and control the loop with loopTools.
Stream tokens in real time with agent.stream.
Validate answers via structuredOutput.

​1. Instantiate an agent

​2. Generate a single response

​Conversation mode

​3. Enable telemetry

​4. Keep exploring

1. Instantiate an agent

2. Generate a single response

Conversation mode

3. Enable telemetry

4. Keep exploring