Create an agent
Create an agent from a template or from scratch, set its instructions, model, actions and knowledge, test it, and deploy it.
You can create an agent in three ways. All of them open the same builder, and all of them start as a draft that nobody can run until you deploy it.
Choose how to start
Open Agents and choose New agent. The page Create an agent offers:
- Start blank: an empty agent that you configure yourself.
- From a template: a ready-made agent that you tailor. Open Browse templates to search the gallery and choose Use template. A template shows what it asks for, which apps it uses and where it asks before it acts.
- Describe it in chat: tell the assistant what you need and it builds the agent with you in a chat on the left, while the configuration fills in on the right.
The builder shows Preview and Build tabs. Use Build with AI to ask for changes in words. Choose Hide chat to give the configuration more room.
Creating an agent needs the builder permission in your workspace. If you cannot create agents, ask your workspace admin. Your plan also limits how many agents you can have.
Configure it
Open the Setup tab and fill in:
Name and description
Give the agent a name and a short description that shows in chat. Add a label to find it later.
Instructions
Under Your instructions, describe what the agent does and how. Type
/to add actions, knowledge, skills and rules into the text.Model
Choose a Primary model. Optionally choose a Secondary model that is used automatically if the primary one fails or is unavailable.
Actions and connections
Choose Add action to give the agent abilities, then connect the apps it needs. See Actions and connections.
Knowledge
Add the documents and knowledge bases it should read. See Knowledge and memory.
Inputs and chat start
Under Inputs & chat start add run input fields, conversation starters and a start message, and choose whether people can attach files.
Other settings include Stop a run after (the maximum duration of a run) and the cost limits under the costs section. A section named What the next run will do reads your settings back to you in plain words.
Test it
In Preview, choose Run test, or ask your agent something. The test uses your current setup, including unsaved changes, and shows a Test badge. If the agent needs an app that is not connected, a notice tells you; you can start anyway, and the agent skips those actions.
Draft or deployed
| State | What it means |
|---|---|
| Draft (Not deployed yet) | The agent is not live. Nothing runs on its own. |
| Deployed (Deployed · release N) | Runs use this version. |
| Changes not deployed | You edited a deployed agent. Runs keep using the deployed version until you deploy. |
Choose Save draft to keep an unfinished agent, or Deploy agent when it is ready. If something is missing, a list in the draft bar tells you what. For a deployed agent, choose Deploy changes, optionally add an update message, and confirm with Deploy version.
Unfinished agents are listed under Draft agents, where you can Resume or Discard them. Up to 5 drafts can be open at once.