Quickstart

This quickstart takes you from an empty workspace to a useful agent you can test in a chat. You can add a workflow or a trigger after the basic behavior is reliable.

1. Describe the job to the Concierge

On Home, describe one real task in the box under What would you like to get done? and choose Start Building. Be specific about the input, the result you want, and what a good answer should contain.

Good starting tasks are bounded and easy to verify, such as summarizing a document, classifying incoming text, or turning structured notes into a report.

The Concierge asks about anything that isn't clear, proposes a plan, and creates the agent once you approve. The agent is private to you until you share it.

2. Check the agent

Open Agents, then the agent the Concierge created, and check that it has:

a name that describes the role;
a short purpose that describes the outcome;
instructions covering behavior, constraints, and output;
a model appropriate for the task;
only the tools the job needs.

Change anything that isn't right. You can also skip the Concierge and build an agent by hand: open Agents, choose Create Agent, and fill in the same details. See create an agent.

3. Add useful knowledge

Open Knowledge, stay on the My Library tab, and upload the documents the agent should use. Keep source material focused: unrelated files make retrieval noisier and make incorrect answers harder to diagnose.

A shared workspace also has a Workspace Library tab for material that members should discover and reuse. It is available even before another member joins. Personal workspaces show only My Library. See manage a workspace for invitations and workspace controls.

What the agent can read depends on who it's shared with. While it's private to you, it can read your My Library and the Workspace Library. Once you share it with the workspace or restrict it to people you add, it reads only the Workspace Library.

A new agent's knowledge access is off. In its editor, set Knowledge access to Everything it can see, or to Selected folders and files for a focused set of sources. See choose knowledge access.

4. Test in a chat

Open Chats, start a conversation with the agent, and send a representative request. Check the answer itself and the run details behind it.

Test at least three cases:

a normal request;
an ambiguous request;
a request the agent should refuse or clarify.

While the agent is working

A message you send while the agent works waits above the message box, in every tab you have open, and goes in when the agent's current turn ends. Until then you can edit or delete it. When several are waiting, they go in one per turn, top first: drag one by its handle to change the order, or focus the handle and use the up and down arrow keys. Every open tab shows the new order.

To change course without waiting, choose Send now (or press ⌘↵, Ctrl+Enter on Windows). The agent reads the message before its next step, and the message appears in the chat marked Steered the agent. Send now carries text only; send files with Send. If the agent finishes before it reads the message, the message goes first in line for the next turn.

Choose Share in the chat's header to preview a snapshot, then Create Share Link to let anyone with the link read it without signing in. Later messages stay out of the snapshot. You can revoke the link from the same dialog. See share a chat for what is included and how to share a newer version.

5. Automate only after it works

When the agent behaves consistently, run it from a schedule or an inbound webhook. If the job has stages that other agents should handle, connect them in a workflow; agent or workflow helps you choose.

Automation multiplies both reliable behavior and bad assumptions. Stabilize the agent first.