Docs Agents

Agents

AI teammates with instructions, skills, connectors and files of their own.

An agent is an AI teammate made for one kind of work: research, copy, QA, sales analysis. It has instructions, skills, tools and files of its own, and works on the tasks you give it.

Making one

Open Agents and press New agent. Say what it should do in a sentence; AI drafts its name, description, instructions and skills, and suggests connectors for the services you named. Prefer to write it yourself? Choose Fill in by hand.

What it's made of

• Instructions: who the agent is, how it works, what it never does.
• Skills: ways of doing one kind of task. The agent sees their names and loads one only when a task calls for it, so it can have many.
• Design system: colours, type, spacing and voice it follows for anything visual.
• Connectors: other services over MCP, like GitHub, Notion or Linear. Each person can connect their own account, so the agent acts as whoever asked.
• Files: its own knowledge. It reads the Space's library too.
• Missions: where it works, everywhere or only on the Missions you pick.
• Model: the AI model it thinks with.

The general agent

Every Space starts with one general-purpose agent. It answers anything, and when you give it a task, it first looks for an agent better suited to it and passes it on, saying why in the task's Activity. It doesn't count against your plan's agents.

What an agent may do

An agent always acts on behalf of the person who gave it the work. It sees what they see and changes only what they may change, so a Guest's request can't reach further than the Guest could.