Using Agents
Agents use LLMs and tools to solve open-ended tasks. They reason about goals, decide which tools to use, retain conversation memory, and iterate internally until the model emits a final answer or an optional stop condition is met. Agents produce structured responses you can render in your UI or process programmatically. Use agents directly or compose them into workflows or agent networks.
📹 Watch: → An introduction to agents, and how they compare to workflows YouTube (7 minutes)Â
Getting started
Mastra model router
Install dependencies
Add the Mastra core package to your project:
npm install @mastra/core
Set your API key
Mastra’s model router auto-detects environment variables for your chosen provider. For OpenAI, set OPENAI_API_KEY
:
OPENAI_API_KEY=<your-api-key>
Mastra supports more than 600 models. Choose from the full list here.
Create an agent
Create an agent by instantiating the Agent
class with system instructions
and a model
:
import { Agent } from "@mastra/core/agent";
export const testAgent = new Agent({
name: "test-agent",
instructions: "You are a helpful assistant.",
model: "openai/gpt-4o-mini"
});
Instruction formats
Instructions define the agent’s behavior, personality, and capabilities. They are system-level prompts that establish the agent’s core identity and expertise.
Instructions can be provided in multiple formats for greater flexibility. The examples below illustrate the supported shapes:
// String (most common)
instructions: "You are a helpful assistant."
// Array of strings
instructions: [
"You are a helpful assistant.",
"Always be polite.",
"Provide detailed answers."
]
// Array of system messages
instructions: [
{ role: "system", content: "You are a helpful assistant." },
{ role: "system", content: "You have expertise in TypeScript." }
]
Provider-specific options
Each model provider also enables a few different options, including prompt caching and configuring reasoning. We provide a providerOptions
flag to manage these. You can set providerOptions
on the instruction level to set different caching strategy per system instruction/prompt.
// With provider-specific options (e.g., caching, reasoning)
instructions: {
role: "system",
content:
"You are an expert code reviewer. Analyze code for bugs, performance issues, and best practices.",
providerOptions: {
openai: { reasoning_effort: "high" }, // OpenAI's reasoning models
anthropic: { cache_control: { type: "ephemeral" } } // Anthropic's prompt caching
}
}
See the Agent reference doc for more information.
Registering an agent
Register your agent in the Mastra instance to make it available throughout your application. Once registered, it can be called from workflows, tools, or other agents, and has access to shared resources such as memory, logging, and observability features:
import { Mastra } from "@mastra/core/mastra";
import { testAgent } from './agents/test-agent';
export const mastra = new Mastra({
// ...
agents: { testAgent },
});
Referencing an agent
You can call agents from workflow steps, tools, the Mastra Client, or the command line. Get a reference by calling .getAgent()
on your mastra
or mastraClient
instance, depending on your setup:
const testAgent = mastra.getAgent("testAgent");
mastra.getAgent()
is preferred over a direct import, since it preserves the Mastra instance configuration (tools registered, telemetry, vector stores configuration for agent memory, etc.)
See Calling agents for more information.
Generating responses
Agents can return results in two ways: generating the full output before returning it or streaming tokens in real time. Choose the approach that fits your use case: generate for short, internal responses or debugging, and stream to deliver pixels to end users as quickly as possible.
Generate
Pass a single string for simple prompts, an array of strings when providing multiple pieces of context, or an array of message objects with role
and content
.
(The role
defines the speaker for each message. Typical roles are user
for human input, assistant
for agent responses, and system
for instructions.)
const response = await testAgent.generate([
{ role: "user", content: "Help me organize my day" },
{ role: "user", content: "My day starts at 9am and finishes at 5.30pm" },
{ role: "user", content: "I take lunch between 12:30 and 13:30" },
{ role: "user", content: "I have meetings Monday to Friday between 10:30 and 11:30" }
]);
console.log(response.text);
See .generate() or .stream() for more information.
Structured output
Agents can return structured, type-safe data by defining the expected output using either Zod or JSON Schema . We recommend Zod for better TypeScript support and developer experience. The parsed result is available on response.object
, allowing you to work directly with validated and typed data.
Using Zod
Define the output
shape using Zod :
import { z } from "zod";
const response = await testAgent.generate(
[
{
role: "system",
content: "Provide a summary and keywords for the following text:"
},
{
role: "user",
content: "Monkey, Ice Cream, Boat"
}
],
{
output: z.object({
summary: z.string(),
keywords: z.array(z.string())
})
}
);
console.log(response.object);
Working with images
Agents can analyze and describe images by processing both the visual content and any text within them. To enable image analysis, pass an object with type: 'image'
and the image URL in the content
array. You can combine image content with text prompts to guide the agent’s analysis.
const response = await testAgent.generate([
{
role: "user",
content: [
{
type: "image",
image: "https://placebear.com/cache/395-205.jpg",
mimeType: "image/jpeg"
},
{
type: "text",
text: "Describe the image in detail, and extract all the text in the image."
}
]
}
]);
console.log(response.text);
Multi-step tool use
Agents can be enhanced with tools, functions that extend their capabilities beyond text generation. Tools allow agents to perform calculations, access external systems, and process data. Agents not only decide whether to call tools they’re given, they determine the parameters that should be given to that tool.
For a detailed guide to creating and configuring tools, see the Tools Overview page.
Using maxSteps
The maxSteps
parameter controls the maximum number of sequential LLM calls an agent can make. Each step includes generating a response, executing any tool calls, and processing the result. Limiting steps helps prevent infinite loops, reduce latency, and control token usage for agents that use tools. The default is 1, but can be increased:
const response = await testAgent.generate("Help me organize my day", {
maxSteps: 5
});
console.log(response.text);
Using onStepFinish
You can monitor the progress of multi-step operations using the onStepFinish
callback. This is useful for debugging or providing progress updates to users.
onStepFinish
is only available when streaming or generating text without structured output.
const response = await testAgent.generate("Help me organize my day", {
onStepFinish: ({ text, toolCalls, toolResults, finishReason, usage }) => {
console.log({ text, toolCalls, toolResults, finishReason, usage });
}
});
Testing agents locally
There are two ways to run and test agents.
Mastra Playground
With the Mastra Dev Server running you can test an agent from the Mastra Playground by visiting http://localhost:4111/agents in your browser.
For more information, see the Local Dev Playground documentation.
Command line
Create an agent response using .generate()
or .stream()
.
import "dotenv/config";
import { mastra } from "./mastra";
const agent = mastra.getAgent("testAgent");
const response = await agent.generate("Help me organize my day");
console.log(response.text);
See .generate() or .stream() for more information.
To test this agent, run the following:
npx tsx src/test-agent.ts