Set up an AI agent

An agent is the thing employees actually talk to. It answers from your knowledge, fills in forms, runs tools, and escalates to a human when it cannot finish the job. Everything you built earlier in this track is what the agent draws on. Time: about 30 minutes, including testing.

Interactive walkthrough
Step-by-step

Open Agents

Click Agents in the left sidebar.

Create the agent

Click New.

Give it a name and a description that says what it handles.

The new agent dialog with name and description fields

Name it something your employees will address in Slack. Teams that give the agent a real name get more traffic than teams that call it “IT Bot”.

Upload an avatar

Images must be PNG, GIF, or WebP for Slack compatibility, under 1MB, and are resized to 512x512.

Add your first rules

The Rules section of an agent with a rule being written

A rule is a natural language instruction that teaches the agent one capability. Use @ to reference the forms, knowledge folders, categories, and tools you set up earlier.Start with two or three, not twenty. For the IT desk built in this track:

When someone reports being locked out of MFA, create a ticket with @MFA reset and set @Category - Account issues. If they say they are fully locked out of every device, set priority to High.

For questions about laptop setup, VPN, or software licences, search @IT runbooks and answer directly. Cite the article you used.

Learn more about writing rules, and read writing agent rules for what belongs in a rule versus the custom prompt

Write escalation instructions

Escalation is what happens when the agent cannot finish. Left alone, it escalates only after knowledge and tools come up empty, creates a ticket, and tells the requester a human is taking over.Override that when your team has its own handoff. Escalation instructions accept the same @ mentions as rules, so you can name the form, category, and status the escalated ticket should carry, which is how a workflow picks the ticket up afterwards.

Before escalating, check @IT runbooks one more time. If still unresolved, create a ticket with @General support and tell the requester that a specialist will follow up within one business day.

Learn more about escalation instructions

Connect knowledge

Select the knowledge folders this agent may search. Agents only reach knowledge you grant them, and the folder’s own channel scoping applies on top.

Learn more about scoping knowledge

Add tools, and decide what needs approval

Tools are what turn the agent from an answering machine into a resolver: looking a user up, resetting a password, adding someone to a group.Every tool in a rule carries an execution policy, set per rule rather than globally:

Auto-execute for read-only lookups.
Requires confirmation for write actions where a “Should I proceed?” prompt is enough. This is the default for creates, updates, and deletes.
Requires approval for anything high impact. The agent buffers the call, opens an approval round, and only runs the tool once the round clears. Approvers default to workspace admins, or you can point it at an approval template.

Keep the lookup on auto-execute and gate the change. That combination is what makes an agent safe to leave running.

Learn more about tool execution policies

If a tool is a Foundry function that takes an API key, pin that input to a Vault credential under Input settings. The agent never sees the value, only a reference is stored, and the rule will not save while a secret input is unpinned.

Learn more about tool input settings

Set personality and Slack behavior

Agent personality settings for tone, length, and emoji use

Tone, response length, and emoji use are structured settings, so use them rather than describing tone in a custom prompt. A custom prompt silently overrides all three.Two settings worth deciding now:

Auto-respond lets the agent answer tickets that arrive by email or from Jira, Linear, and GitHub, with an optional delay so context can land first.
Conversational form filling collects short forms one question at a time in chat instead of showing a form. It applies to forms at or under the field threshold, three by default, and never to private forms.

Learn more about personality, auto-respond, and conversational form filling

Test it before deploying

Use the chat panel on the agent page. A yellow Testing Mode banner confirms the session is sandboxed: the agent matches rules, calls tools, and picks responses for real, but nothing is written and nothing posts to Slack or Teams.Expand the Debug disclosure under each response to see which rules matched, which tools ran and with what inputs, and how the agent classified its own answer. That is how you find out a rule is not firing, rather than guessing from a bad reply in production.

Learn more about Testing Mode

Deploy to a channel

The Connections section listing channels the agent is deployed to

In Connections, add the channel you set up earlier. One agent per channel, and you can add or remove channels at any time.

Learn more about deploying agents

Agents and workflows do different jobs

An agent handles the conversation. A workflow handles the deterministic sequence. They do not call each other directly: the agent creates or updates a ticket, and the ticket’s state is what a workflow triggers on. That is why escalation instructions let you set a form, category, and status. Those are the handoff.

Learn more about agents compared with workflows

What to watch in week one

Open Chat logs on the agent page. Every conversation is recorded with the result the agent assigned itself, the requester, the channel, and any thumbs up or down.

A high Missing Knowledge rate means the gap is in your documentation, not the agent.
A high escalation rate on questions you know are documented usually means a rule is missing or too narrow.
Frequent requests for more information often means a form is asking for things the agent could have inferred.

Learn more about chat logs and debugging patterns

Everything is built. Now open it up.

Go live

Run the pre-launch checks, announce it to your team, and know what to look at on day two.

Start here

Playbooks

How to

Best practices

Resources

Agents and workflows do different jobs

What to watch in week one

Next

Go live

​Agents and workflows do different jobs

​What to watch in week one

​Next

Go live

Agents and workflows do different jobs

What to watch in week one

Next