Skip to main content
A custom agent is a reusable worker you define once and call by name. It has its own system prompt, its own tool list, optionally its own model, and optionally a pipeline of steps. Manage them at Agents in the sidebar, or with /agents. The assistant can also create, change or delete an agent for you from a chat. It always asks first, and the request shows the tools, connectors and model the agent would get as well as its prompt.

​
What you configure

The tools you can grant are web search, fetching a URL, searching your indexed sites, creating artifacts, generating PDFs, generating images, using connectors, and using integrations.

​
Steps

An agent can run as a pipeline rather than a single pass. Each step has its own prompt, model and tools, and is either a generate step that produces something or a validate step that judges it. A validate step returns a pass or fail with a reason, and you choose what happens on failure: stop, continue anyway, or retry up to five times. Twelve steps maximum.

​
Using one

Attach an agent to a chat from the composer and the assistant prefers it for matching work. Or let the assistant pick: it sees your saved agents and their descriptions and calls the one that fits. A good description matters more than a good name here, because the description is what the assistant reads when deciding.

​
Limits