Deployments and Channels is how you take an agent live and choose where people reach it — a deployment pairs a version of your agent with a channel, a schedule, or Copilot.
Overview
A deployment makes an agent available to use — it pairs a version of your agent (or team) with a place people reach it. Each agent has a Deployments tab listing its deployments, each AI team has the same tab, and there's a workspace-wide deployments list spanning every agent. All three surfaces use the same deployment types, form, and list actions; the tab appears only if your role grants deploy access, and never for External agents.
You create a deployment with New Deployment, give it a name and notes, pick the agent or team to deploy, choose a deployment type, and pick which version to release. A deployment can be Live, Paused, or left as a Draft. The deployments list also has Created On and Last Updated By columns, hidden by default — turn them on from the column-visibility control, or hover the Last Updated date for the same detail as a tooltip.
Deployment Types
When you create a deployment you choose Deploy On — one of three types:
Type | Description |
|---|---|
Channels | "Deploy on Slack, MS Teams and other communication channels." Put the agent into a messaging channel. |
Schedule | "Automate your agent to work at predefined intervals." Run the agent on a timer rather than in response to a person. |
Copilot | "Interact with your agent using Copilot." Make the agent available through the platform's Copilot (one Copilot deployment per agent). |
Pick Channels to put the agent where users already chat, Schedule for unattended runs, and Copilot for the built-in assistant experience.
Deploying an Agent to a Channel
Choose Channels, then walk the channel setup:
Select App — pick the channel app (such as Slack or Microsoft Teams). The available channels come from your connected channel apps.
Select Action — pick the action for that app.
Select Connection — choose the connection that authenticates it.
Inputs — fill any inputs the channel needs.
Setup — confirm the agent settings and finish.
Select Create to save the deployment. A new deployment starts as Paused — switch it to Live with the toggle on its row in the deployments list. The exact list of channels depends on which channel connectors your workspace has.
What Happens Behind the Scenes
A channel deployment creates a small dedicated listener for that channel: the platform takes a ready-made recipe for the chosen channel (for example "AI Agent (Slack Deployment)"), wires in the connection you selected and the agent being deployed, fills in your setup inputs, and installs it. That listener is what receives each incoming message, files it into the right conversation, runs the agent, and hands the answer to a channel-specific reply step.
It starts paused — the deployment isn't answering anyone until you set it Live. Deleting the deployment removes its listener with it, so nothing keeps running in the background.
How Each Channel Identifies the User and Replies
Channel | Behavior |
|---|---|
Slack | Listens for new messages; messages written by bots and deleted-message events are ignored, so the agent never answers itself. The sender's Slack profile email is looked up and becomes the user's identity, so permissions and knowledge access follow that person. Replies post back into the channel — optionally into the message's thread. |
Microsoft Teams | Connects through an Azure Bot; the agent replies into the same Teams conversation the message came from, and shows a typing indicator while it works. |
Replies go straight back to the sender's WhatsApp number. | |
Gmail | The agent answers by replying in the same email thread, so the exchange reads as a normal email conversation. |
Outlook | The agent sends a reply email ("Re: …") back to the sender. |
Voice (Twilio, Ozonetel) | The incoming call is handed to the platform's voicebot for a spoken conversation. |
Slack-Specific Options
Slack setup includes three behavior toggles:
Toggle | Effect |
|---|---|
Reply in thread | Answers go into the message's thread instead of the channel, and follow-ups in that thread stay in the same conversation. |
Show thoughts as emoji | While the agent works it adds a loading reaction to your message, then swaps it for a ✅ when the answer is posted, instead of posting intermediate "thinking" messages. |
Delete chain of thoughts | Cleans up the agent's intermediate thought messages once the final answer is posted. |
How Channels Format the Agent's Answers
The reply step adapts the answer to what each channel can display:
Slack — the answer is converted to Slack's message formatting: citation markers become readable references, leftover HTML is stripped, and markdown tables are reformatted so they render properly. File attachments arrive as a File button, and thought updates (if enabled) appear as 💭 attachment lines.
Microsoft Teams — the answer is rebuilt as an Adaptive Card: text, images, and citations (with source-app icons) become card elements, and tables become card tables (very long tables are capped at the first 100 rows).
WhatsApp — plain text messages; images and videos in the answer are sent as real media messages with the text as a caption. When the agent offers choices, up to three become tappable reply buttons; more than three become a tap-to-open options list.
Gmail — the reply is sent as an HTML email inside the original thread.
Outlook — the reply is sent as a plain-text email.
Note: Rich in-chat widgets (charts, interactive blocks) are a Copilot capability — channel replies stick to what the channel itself supports.
Running an Agent on a Schedule
Choose Schedule and configure two things: a prompt for the agent (what it should do on each run) and a schedule (Setup Schedule — an interval or a custom schedule). The agent then runs automatically at the times you set, without a person prompting it. Use this for recurring jobs like a daily summary or a periodic check.
Each run works exactly like a user sending the configured prompt to the agent: the prompt is recorded as the incoming message and the agent's output as the reply, so scheduled runs show up in Sessions and Traces like any other conversation. Consecutive runs continue the same ongoing session rather than opening a new one each time.
Calling the Agent Over an API
Agents can be invoked over a REST API. A request carries the agent id, the message, and optionally: a sessionId (pass the same id again to continue that conversation; omit it to start fresh), a userEmailAddress (so the run uses that person's identity and permissions), attachments, and a flag to include citations. The response returns the agent's answer — text plus any offered choices — along with citations and the session id for follow-ups. A streaming variant delivers the answer as it's generated. API conversations appear in observability with API as their channel.
Deployment Status and Pausing
Each deployment shows a status: Live (deployed and serving), Paused (saved but not serving — new deployments start here), or Draft (an older deployment that was never set Live or Paused). You switch a deployment between Live and Paused from its row in the deployments list — there's no toggle inside the form. Under the hood, Live/Paused simply resumes or suspends the deployment's channel listener, so pausing takes effect immediately and nothing is lost — set it Live again to bring the agent back on that channel.
Which Version Gets Deployed
A deployment uses Full Deployment — "Release a single version of your agent to all users at once" — and you pick the Version to release. So deploying always publishes one specific, chosen version to everyone on that channel. Publish a new version when you want changes to go live, then update the deployment to that version.
The Copilot deployment is the exception: when you publish with the "deploy on Copilot" option checked, the platform creates the Copilot deployment if it doesn't exist and moves it to the freshly published version automatically — you don't update it by hand.