Skip to main content
An agent is a named assistant that you define once. It has an instruction, tools, optional knowledge and its own conversation starters. Members chat with it, flows run it in an Agent step, and jobs start it from an event or a schedule. Agents belong to the instance, not to an automation workspace. Guests cannot use agents. Creating an agent from a description needs an AI model on your instance. See AI configuration.

​
Where to find agents

​
Create an agent

1

Open the agents page

In Studio, select Agents in the sidebar.
2

Describe the job

Write what the agent should do and press Enter. Press Shift + Enter for a new line. A description can have up to 2,000 characters. Four suggestions under the box fill it for you: Task triage, Ticket writer, Meeting follow-up and Weekly status.
3

Check the result

The agent opens in the agent editor. Read the instruction, then change what does not fit.
Select Create manually instead to start with an empty agent named New agent. An agent made from a description gets a name, a description, an instruction, up to three skills and tools of Collabase AI that only read. It never gets a tool that changes data, knowledge, conversation starters or an identity of its own. Add those in the editor when you want them. If your instance has no AI model, the message says that Collabase AI is not set up. Each person can create up to 100 agents.

​
Who sees and manages an agent

Select Share at the top right of an agent to set who sees it. Invited people and groups get one of two roles. The creator and system administrators always manage an agent. A line under the name shows who sees the agent and who manages it. A person who only uses an agent sees the editor without the Jobs card and cannot change anything in it.
A flow can run a restricted agent only while the person who published the flow may still see the agent. A flow that loses sight of its agent stops with a message instead of widening who can use it.

​
Use an agent

The model of an agent is set at the top of the editor. Automatic uses the model of the instance. The other choice uses the provider that was set for the automation workspace that runs the flow or job. A chat always uses the model of the instance.

​
Usage and token limit

Select Usage at the top of an agent to see the last 30 days. Managers can enter a Monthly token limit. Once an agent has used that many tokens in a month, new runs of its flows and jobs stop, and a job shows that it is paused. Leave the field empty for no limit. The check happens before a run starts, so the run that crosses the limit may finish. Chats and the builder are not counted.

​
Good to know

  • An agent saves itself about a second after your last change. The top of the editor shows Saving…, Saved or Saving failed. Without a name and an instruction an agent is not saved.
  • There is no delete button on the agent page yet.
  • Changes to an agent, its identity, its sharing and its token limit appear in the audit log as Automation agent created, changed and deleted. See Audit events.

​
Next steps

The agent editor

Instruction, tools, knowledge and conversation starters.

Identity and jobs

Give an agent an account of its own and let events start it.

Agent step

Run an agent inside a flow.

MCP servers

Connect servers whose tools every chat can use.
Last modified on October 27, 2026