Operator

Operator is the chat window inside Horizon where you talk to your deployed agents, have the Operator agent hand work to them, get a good-morning briefing, and open an agent from Meetings or Sites.

Operator is a chat window inside Horizon where you talk to your organization’s deployed agents without leaving the page you are on.

Open Operator

Select Ask AI, the sparkles button in the bar at the top of Horizon, next to notifications. Operator opens as a floating window in the bottom-right corner, over whatever page you have open.

The window has two layouts. Compact shows one view at a time: your chats, the agent picker, or a conversation. Expanded shows your chats on the left and the open conversation on the right, and you can resize it.

If there is no Ask AI button, Operator chat is turned off for your organization. See Turn Operator chat on or off.

Which agents appear in Operator

Operator lists every active agent deployment in your organization that is allowed in Operator chat. A deployment is allowed when one of these is set:

Deploying one of your own agents: choose Operator only (no external channel), or choose Slack or Microsoft Teams and tick Also allow in Operator chat.
Deploying an agent from the Store: turn on Operator Chat under Communication Channels.
An existing deployment: turn on Allow in Operator on the deployment’s Configuration tab. The person who deployed the agent, or an owner or admin, can change it.

Only active deployments appear, so a paused agent drops out of the list until you resume it. See Deploying agents for the deployment states.

The list always starts with Operator, Horizon’s own agent for your organization. It sets up your company profile, goals and offerings with you, recommends agents to deploy, and helps you connect your tools. It can also hand work to your other agents and sends the good-morning briefing.

Start a conversation

Open Operator and select New chat, the + button above your chats.
Pick an agent. The search box matches agent names and descriptions.
Type your message and send it.

The agent answers using its own instructions, skills and connections. Once the conversation has used Horizon Tokens, its header shows how many.

Attach files

Attach files with the paperclip (Attach files) button, by dragging them onto the window, or by pasting them into the message box. You can attach images, PDF, plain text, CSV, Excel (.xlsx) and Word (.docx) files of up to 10 MB each.

The agent sees JPEG, PNG, GIF and WebP images.
The agent reads the text of PDF, plain text, CSV, Excel and Word files. Long files are cut down before the agent reads them.
The agent reads up to 10 attachments per message.
An iPhone HEIC photo is converted to JPEG in your browser before it is sent. If the photo cannot be read, Horizon tells you and asks for a JPEG instead.

Continue a conversation

Your chat list shows your own Operator conversations, most recently active first, with the agent’s name, the latest message and when it was sent. A dot marks a conversation with replies you have not read. The search box above the list filters it by agent name.

Select a conversation to reopen it and carry on. Your conversations are your own: other people in your organization do not see them in their Operator.

Ask the Operator to hand work to an agent

In a chat with the Operator agent, ask for something one of your other agents does, such as finding prospects or drafting emails. The Operator hands the work to that agent and brings its result back into your chat.

Which agents it asks: only agents you could pick in Operator yourself. If none of them does the job, it recommends an agent from the Store instead.
Before it asks: it tells you which agent it will ask and what it will ask it to do, and waits for you to agree.
How the agent works: on its own, with its own instructions, skills, connections and approval rules. The agent does not see your conversation, so the Operator puts everything it needs into the request.
More than one agent: the Operator can split a big request between agents, for example one agent finds prospects and another drafts emails to them. It asks at most three agents per message.
How long it takes: an agent can take a few minutes. An agent still working after five minutes is stopped; steps it finished before then may have run, and the Operator tells you what happened.

When the agent finishes, the Operator tells you what it reported.

What it costs

Asking an agent costs the same as using that agent directly, on top of your message to the Operator:

your message to the Operator, at the Operator’s price per run;
the asked agent’s own price per run;
the agent’s skills, each at its usual price every time it runs.

No agent costs less when the Operator asks it. If you approve an action the agent stopped on (see below), the agent’s finishing run and the Operator’s report back are each charged once more at their price per run. See Models and Run Cost for how a price per run works.

When the agent needs your approval

If the agent reaches an action that needs approval, it stops there and the action is not run. Anything it finished before that point was done. The approval card goes only to you, the person who asked, in your chat with the Operator:

Approve it, and the agent carries on from that action. The Operator then tells you what happened. If the agent reaches another action that needs approval, you get another card.
Decline it, or let it expire, and the action is not done. The Operator adds a line to your chat saying so.

When an agent stops for your approval, the Operator stops there too, and no other agent starts in that message. Ask for the rest once you have decided.

Only you can approve it. If the agent’s skill names its own approvers and you are not one of them, the action is refused and not run.

Your good-morning briefing

On weekday mornings the Operator agent posts a briefing into a conversation with you. Each one starts with “Good morning”, and every briefing goes into the same conversation, so earlier ones stay there to scroll back through. Reply in that conversation to ask the Operator about anything in it.

When: from 07:00, Monday to Friday, in your time zone, the one under Display settings in your user menu. If no time zone is set, UTC is used. The briefing arrives shortly after 07:00. If it cannot be sent before noon, none is sent that day. You get at most one a day.
Who gets it: owners and admins receive it by default; other members do not. Anyone can turn their own briefing on or off with Send me the good-morning briefing under Display settings in the user menu. It is sent only while Operator chat is on for your organization.
What it costs: nothing. Horizon builds the briefing from your organization’s data, without the model, so it uses no Horizon Tokens. A reply to it is an ordinary message to the Operator.

A briefing has up to three parts:

What happened since your last briefing: new contacts and leads, deals that moved stage, customer payments received, invoices that went overdue, and meetings held. A first briefing looks back 24 hours, and no briefing looks back more than four days, so Monday’s covers the weekend.
What needs you: approvals waiting on you, and your own tasks that are overdue or due today.
What your agents did: each agent’s finished and failed runs, runs held waiting for tokens, yesterday’s token spend and your balance. It warns you when the balance is empty or would last less than three days at yesterday’s rate.

The briefing shows only what you can see in Horizon, so a module you cannot open adds nothing. A part with nothing to report is left out, and on a quiet day the briefing says there is nothing new. If a part cannot be loaded, the briefing says so instead of reporting a quiet day.

Open Operator from a module

Some modules have a button that opens Operator with the right agent already chosen and the record you were looking at passed to it as context:

Meetings: on a meeting’s page, select Ask about this meeting. The Meeting Agent opens with that meeting as context and suggested questions such as “What did we decide?”
Sites: select Build a site with an agent on the Sites list, or Edit with agent on a site’s page. You describe the site you want, and the agent builds or changes it.

Operator opens on a start screen showing the agent, a few suggested questions and a Back link to the page you came from. Nothing runs, and no Horizon Tokens are used, until you send a message or pick a suggestion. From then on the conversation is in your chat list like any other.

Meetings and Sites are available only to organizations that have those modules. Their buttons need Operator chat to be on, and the module’s agent to be deployed and allowed in Operator.

Turn Operator chat on or off

Operator chat is on for new organizations. An owner or admin can turn it off, or back on, with the Operator Chat switch on the General tab of Settings.

While it is off:

The Ask AI button and the Operator window are hidden for everyone in the organization.
The Operator Chat switch on the Store deploy page is unavailable, with a link to Settings.
The Meetings and Sites buttons cannot open a conversation.
No good-morning briefings are sent.

Turning it off does not delete any conversations or change which agents are allowed in Operator. An API key with the Operator Chat permission can read and change the same setting; see API Key Management.