Guide · Collaboration
Multiplayer AI: how humans and the agent work one conversation together
Handoff is a relay race: the bot runs, then stops, then a human runs. Multiplayer AI is a team on the field at once. The AI keeps helping while a human leads, teammates can see each other, and anyone can pull the AI in for a draft. Here's the model and the mechanics.
The problem with handoff
Almost every AI support tool models a conversation as belonging to one responder at a time. The bot owns it, then it hands off, then a human owns it. The instant a person steps in, the AI goes dark: it stops answering, and often the best it can do afterward is post a canned “a teammate will be with you shortly.”
That throws away the thing the AI is best at right when a human could use it most. The agent who just picked up a live chat still has to look up the refund window, recall the exact plan limits, or draft a careful reply to an upset customer, alone, while the visitor watches the typing indicator sit still.
The multiplayer model
Multiplayer AI treats a conversation as a shared room rather than a baton. A human can lead while the AI stays available as a colleague you direct, and teammates can be present alongside you. Three capabilities make that real:
- Presence. When you open a conversation, everyone else viewing it sees an “Also here” chip with your name, and you see theirs. It updates the moment someone opens or closes the chat, so two agents don't unknowingly answer the same visitor.
- Live typing. The visitor, your teammates, and the AI all surface a typing indicator in real time. When the AI is preparing a draft you asked for, you see an “AI is drafting” cue, staff-side only.
- Summon the AI. While you lead the thread, you can pull the AI in for a draft reply whenever you want, without handing the whole conversation back to it.
Summoning the AI: @ai and Ask AI
There are two ways to ask the AI for help mid-conversation. Type @ai followed by an instruction in the composer, or click the Ask AI button. With @ai you steer it (“@ai push back politely on the discount request”); with no instruction it simply drafts the next reply from the conversation so far.
The draft is generated exactly like an autonomous reply: retrieval over your knowledge base, the same tools the agent is allowed to call, the same model your workspace is configured to use, and the same PII redaction applied to the visitor's messages before anything reaches the model. What you get is a production-quality reply, just held back for a human to look at.
Draft by default, send when you decide
A summoned reply always arrives as a draft, never straight to the visitor. You then choose:
- Send as AI. Post the draft unchanged into the thread as the AI persona, in one click. The visitor sees it as a message from the assistant.
- Edit and send as me. Move the draft into your composer, adjust the wording, and send it under your own name.
- Dismiss. Throw it away and write your own.
This is the human-in-the-loop lever most teams want: the speed of an AI draft with a person deciding what actually reaches the customer.
What stays staff-side
Everything about the collaboration layer is invisible to the visitor. The presence roster, the live typing between agents, your @ai instructions, and the draft before you send it are all fanned out to an operators-only channel scoped to that one conversation. A visitor never learns which staff are reading their chat, and never sees a draft you didn't send. The only thing they ever see is a message you deliberately post.
When to reach for it
Multiplayer AI shines in the messy middle of a conversation:
- A human has taken over a sensitive thread but wants the AI to pull the exact policy language so they don't misquote it.
- A newer agent wants a first-draft answer to react to, rather than starting from a blank box.
- A senior agent quietly joins a live chat to back up a colleague, and both can see they're there.
- The AI is handling a chat solo, a human notices it's about to get something subtle wrong, and steps in without the AI going silent for the rest of the conversation.
The point isn't to replace autonomous AI or full human handling. It's that the boundary between them stops being a cliff. For the wider argument on where that line should sit, read the ROI of AI agents vs humans. For the foundations of what the agent can do on its own, start with what an AI support agent is.
Quick FAQ
What is Multiplayer AI?
It's a mode where a human agent, the AI agent, and any other operators can be live in the same conversation at once, instead of the conversation belonging to exactly one of them. You get presence (who else has this chat open), live typing indicators, and the ability to summon the AI for a draft reply while a human leads the thread.
How do I ask the AI for a reply while I'm handling a chat?
Two ways. Type @ai followed by an instruction in the composer (for example, '@ai summarise our refund window'), or click the Ask AI button. Either produces a draft built from the same knowledge base, tools, and model the AI would use on its own. You can edit the draft and send it as yourself, post it unchanged as the AI in one click, or dismiss it.
Does the visitor see me typing @ai, or the draft before I send it?
No. Your @ai instruction and the draft are staff-side only, fanned out to an operators-only channel for that conversation. The visitor sees nothing until you actually send a message. The 'who's here' roster is also operator-only, so a visitor never learns which staff are reading their chat.
Is the AI draft as good as an autonomous reply?
It's built the same way: the same retrieval over your knowledge base, the same tools, the same model, and the same PII redaction on the visitor's messages before anything is sent to the model. The only difference is that a human reviews it before it reaches the visitor.
What stops two agents from replying over each other?
When another operator has the conversation open, you see an 'Also here' chip with their name, plus their typing indicator in real time. It's a coordination signal, not a hard lock: support teams told us a lock is more annoying than a visible presence cue that lets them sort it out.
Try it on your own LLM keys for $1.
Founder rate. Locked in for life if you join while we're inside the first 1,000 sites. No card on file, just paste your model key and your widget is live.