A chat model answers a question. An agent goes further: it can look things up in your tables, run your flows, search your documents, use other apps, and take several steps to finish a job, all while staying inside the rules you set. Examples:
  • A support assistant that finds a customer’s orders, answers from your help articles and, with a person’s approval, issues a refund.
  • A sales helper that reads your products, looks up a customer in Stripe and drafts a reply.
  • A research assistant that searches the web and summarises what it found into a new page in Notion.

​
What an agent is made of

​
How it behaves

  • It acts as the person who asked. When a signed-in user runs the agent, its tools follow that person’s access rules. If the user cannot read an order, the agent cannot read it for them. You can switch this off for agents that run on behalf of your own servers.
  • It asks before it changes things. Reading is free; adding, changing or deleting waits for a person to approve, unless you tell a tool never to ask.
  • It is cautious with the outside world. After it reads something it did not get from you, such as a web page, anything it changes afterwards waits for a person. This protects you from a page that tries to tell the agent what to do.
  • It stays on budget. Every run has limits. When one is hit, the run stops and says so.
  • Everything is recorded. Each run keeps its question, steps, tools, answer, tokens and cost.

​
Run it from anywhere

​
Before you start

You need a chat model (see Providers and models). To let it search documents you also need an embedding model.

Create your first agent

From a blank page to a working assistant in Studio.

Tools

Everything you can give an agent.

Approvals and limits

Keep it safe and on budget.

Run from your app

Run, stream and approve from TypeScript or Python.