Skip to main content
An assistant talks to your customers. An agent works for you. You describe a job, like “every morning, read yesterday’s conversations and post a short summary to the team”, and the agent does it using the apps you connect. Every time an agent starts work is a run. Each run is a chat you can open, read and join.

​
When to use an agent

  • Recurring work: a daily briefing, a weekly report, chasing unpaid invoices.
  • Work started by another app: a new order or a form reply comes in and the agent handles it.
  • Questions from your team: mention the agent in a comment and it answers in the same thread.
To decide when an agent works by itself, add Routines.

​
Create an agent

1

Open Agents

In the sidebar, click Agents, then Create Agent.
2

Name it and pick a template

Give it a Name and pick a Template, the job to start from. You can change everything later. Prefer a guided setup? Click Set up step by step.
3

Write the instructions

On the agent’s Settings tab, fill in Instructions. Tell your agent what its job is and how to do it, as you would tell a teammate. Every run follows these instructions.
4

Test it

Click Test Agent to try it in a draft conversation. Test runs are not saved to the Runs list.
5

Decide when it runs

Open the Routines tab and add a schedule, Autopilot or a webhook. See Routines. You can also start a run any time with Run.

​
The agent’s tabs

​
Settings

  • Model: the model the agent thinks with on every run.
  • Instructions: the job, in your own words.
  • Tools: Web search lets the agent search and read the web. Computer gives it its own computer to run code and work with files, and Internet sets what that computer can reach (Full, Trusted sites or Off). Run other agents lets it start, watch and stop other agents. Memory lets it keep notes about its job.
  • Approvals: turn on Ask before changes to approve changes before the agent makes them. Routine runs never wait, since nobody is there to answer.

​
Actions

You don’t have to add apps one by one. On every run the agent can use every app you connect, like Gmail or Google Calendar, and asks you to connect an app the first time it needs one. Add an app on the Actions tab only to lock its settings (for example, which sheet it writes to) or to set up one the agent cannot find on its own, like your own API.

​
Runs

Each run shows a status: Running, Completed, Waiting approval, Waiting response, Needs a decision, Needs a connection, Failed or Canceled. Open a run to see what the agent did, or write in it to give more directions. When a run needs you, it also shows up in Tasks.

​
Mention an agent in a comment

Type @ and the agent’s name in a comment on a contact or a table record. The agent starts a run and answers in the same thread.

​
Memory

What an agent learns about its job lives on the agent. See or delete it on the agent’s Memory tab. Memory explains how this differs from contact and user memory.

​
FAQ

An assistant replies to your customers on your channels. An agent works for your team, doing a job you describe, on a schedule or whenever you ask.
Check that the agent has Instructions. Every run follows them, so without instructions the agent has nothing to do.
The agent needs an app you haven’t connected yet. Open the run, or Tasks, and connect it. The run continues once it is connected.
Its routines stop and it disappears from the list. Past runs stay in your chats.
Yes. Open the agent’s options and duplicate it. The copy is named after the original with “(Copy)”.