An agent is a member of the workspace, so work reaches it the same way it reaches anyone.
Set the agent as the assignee on a record. It picks the work up and acts according to its persona — drafting a reply, categorising, summarising, answering.
Useful when work arrives continuously and the first step is always the same: an inbox of requests, incoming form submissions, first-line questions.
Add an Agent step to a workflow: when a record is created, have the agent read it and fill in a summary, a category, or a draft response. The step runs the agent on each item or once for all of them. See The Agent step in workflows.
This is the higher-volume pattern — it runs without anyone assigning anything.
Decide deliberately what the agent finishes and what it only prepares.
The usual pattern is agent drafts, person approves, and the approval step is one click on a record.
An agent asks before it acts in a connected tool, such as sending a Slack message or an email, and before it uses someone's personal connection. Changes inside the workspace run without asking, except that updating more than 20 rows by a condition shows a preview first, with Update N rows and Decline.
When an agent working on an assigned record reaches an action that needs approval, it pauses:
Approve lets the agent carry on. Deny, or no answer within a day, returns the record to the status the agent picks work up from, and nothing is sent. Unassigning the agent drops the request and nothing is sent; the record keeps its current status.
In a chat, the same check is an Approval required card with Approve, Deny and Always allow in this conversation. Always allow covers every later action in a connected tool for the rest of that conversation, not only the one on the card, so don't use it in a conversation that may message a client. For a custom agent, set Integrations to Read & write in the agent's capabilities and write actions ask for approval in chat before they run.
When an agent creates or updates many rows at once, the whole change is one batch with one Undo on that step in the chat, kept for 7 days. You can also ask the agent to undo it. Undo works field by field: a value someone changed after the batch keeps the newer value, and the step says how many rows had one. Only the person who ran the change, or the org owner, can undo it, and it does not take back what a workflow already did when those rows changed, such as an email it sent. A page the agent edited has an Undo for 30 seconds, and none when the page is very large. A single-record edit has no Undo button and shows in the record's history. A message the agent sent cannot be recalled.
The first time an agent wants to use one of your personal connections, such as your mailbox, it asks in your own chat with a Use your connection? card: Allow this agent, Allow all agents or Not now. Sending still asks you first, unless you already clicked Always allow in this conversation in that chat. To review or withdraw a grant, go to Profile › Settings › Integrations, open the connection, and click Revoke in its Agent access section.
An agent's instructions should say when to stop and get a person, and that path should exist — assigning to a human, changing a status, posting in a channel. An agent that answers everything confidently is a liability; one that escalates well is a colleague.
Agent actions appear as that agent, in the record's history and comments. You can always see what it changed and when — the same audit trail as a person. A bulk change also carries an Undo on its step in the chat for 7 days.
Related: What is an agent? · Create an agent and give it knowledge · Automate the repetitive parts