# Multiplayer sessions in Slack

> How more than one person shares a single agent session in a Slack thread — and the two owner gates that keep it safe: who may take part, and which actions run.

Source: https://openap.org/docs/multiplayer-sessions

An OAP agent session in Slack is a **thread**, and a thread can hold more than one human.
When several people are in a thread with an agent, they share one session — one transcript, one set of
participants, one authorization boundary. Two owner gates keep that shared session safe: one decides
 **who may take part**, the other decides **which actions run**.

[Video: A multiplayer session with two owner gates: approving Alex to take part, then approving the production deploy.](https://openap.org/media/multiplayer-approval.mp4)

## A new participant is a decision

Anyone can open a thread, but a new voice acting in a session is a permission decision, not a free-for-all.
In the clip, Jordan starts the rollout and Alex asks to help drive it. Instead of acting on Alex's message,
`srebot` pauses and asks an **owner** to approve Alex's participation first.

![A new participant is gated: srebot waits for an owner to approve Alex before acting on his message.](https://openap.org/media/mp-join.png)

> **Security: A new participant is gated, not assumed**
>
> Alex's request is held until an owner approves it. Only after Jordan approves is Alex's message picked up
> and acted on. The same owner check that guards an action guards who is allowed to steer the session in the
> first place.

## Sensitive actions need the same sign-off

With Alex in, `srebot` works toward the deploy — and hits a step too consequential to run unattended:
shipping `hotfix-1.4.2` to production. It posts an **approval card** and pauses, setting its status to
_waiting for an owner to approve_. Because owners are easy to miss in a busy thread, OAP also **DMs the card
to the owners**, so any of them can approve without opening the channel.

![The production-deploy approval, DMed to an owner so they can approve without opening the channel.](https://openap.org/media/mp-dm.png)

The card is a first-class **interaction**, not free-form text — the buttons, the audit trail, and the
"who may click this" rules are all built in rather than improvised per message.

> **Security: Any owner may approve — and quorum is one**
>
> At click time, OAP checks that the person who clicked is an **owner** of the resource the gate protects. Any
> one owner is enough — approval is quorum-of-one — but a bystander clicking **Approve** is rejected. The
> check happens when the button is pressed, against live authorization state, not when the card was shown.

## Approval resolves everywhere at once

Because an approval is one interaction with one outcome, resolving it in the DM resolves it **in the thread
too**. The card's buttons are replaced in place with a green _Approved by Jordan_ line, `srebot` posts that it
received the approval, and the rollout runs to completion — all visible to everyone still watching the thread.

![Once an owner approves, the card resolves in place and srebot completes the rollout.](https://openap.org/media/mp-complete.png)

That is the multiplayer property end to end: many humans, one shared session, gated at both the door and the
dangerous step, resolving consistently for everyone at once.
