Multiplayer AI coding, defined term by term

This glossary defines the terms used across the multiplayer AI coding category and in Poly (usepoly.co), one or two sentences each, so that a page, a demo or a security review can point at a definition instead of restating it. Category terms come first, then the terms Poly uses for its own parts.

Category terms

Multiplayer AI coding workspace

A place where several people work with AI coding agents on one codebase and can see each other doing it. It comes in three shapes: several people steering one live session, each person running their own agent thread on a shared project, or one agent living in a team chat. Full definition.

Multiplayer agentic coding

The first of those shapes: a team steering one AI coding agent together, live, in one shared session, with every prompt, reply and change visible to everyone and attributed to the person who caused it. Full definition.

Shared agent session

One running agent that more than one human can prompt, question, approve or stop. The opposite of a private session, where the person who opened the terminal is the only one who can act.

Coding agent

An AI system that reads and edits code and runs commands in a workspace to complete a task, rather than only suggesting text. Claude Code and Codex are the two Poly runs.

Agent harness

The program that wraps a model into a coding agent: the tools it can call, the loop it runs, the permissions it asks for. Claude Code, Codex, OpenCode and Cursor's agent are harnesses; a workspace can run one or several.

Multi-agent

Several agents working for one person, usually in parallel on split tasks. More agents, not more people; often confused with multiplayer.

Approval before action

A gate where a human can allow or refuse a specific change or command before the agent performs it, as opposed to reviewing the result afterwards in a diff or a pull request.

Cost attribution

Recording, per unit of agent work, which person caused it and what it cost. Impossible in products where each user's spend sits on their private subscription.

Sandbox

An isolated environment the agent runs in, with its own filesystem view, limits on resources, and no access to the host or to secrets it was not given. What makes one secure.

Egress allowlist

The list of network destinations an agent's sandbox may reach; everything else is blocked. The alternative, open internet from the sandbox, lets a bad instruction send a project's files anywhere.

Poly terms

Room

Poly's unit of collaboration: a workspace with a git history, a timeline, a group chat and one shared agent, with an invite-only member list. A project and a room are the same thing.

Turn

One run of the agent in response to one prompt. Every turn starts in a fresh sandbox, ends as a git commit, and carries a receipt. Turns in a room run one at a time.

Receipt

The line at the end of every turn saying who ran it, which model, how long it took and what it cost in credits. It is the record cost attribution needs.

Prompt queue

Where prompts wait, visibly and in order, while a turn runs. Anyone can see whose prompt is next and remove their own.

Working mode

A room's rule for how much the agent may do without asking. Full control acts freely; ask before big steps stops at bulk changes, network access, deletions and commands; ask about everything stops at every change; plan only proposes and never edits. Changing the mode is recorded in the timeline, and an owner can lock it.

Approval card

The plain-language card that appears on every member's screen when the agent proposes something the working mode requires a decision on. Any member decides; the timeline records who. Unanswered cards are denied after thirty minutes.

Review mode

An optional room setting under which a prompt itself waits for a teammate's approval before it runs.

Revert

Undoing one turn in a single click. It is a new commit that reverses that turn's changes, not a rewrite of history, and it is itself a recorded turn.

Group chat

The room's conversation among people, hidden from the agent by design, so that arguing about the plan does not become the prompt. Threads and polls live there too.

Poll

A vote in the group chat, visible to members only, used to settle a fork in the road before anyone prompts.

Decision log

The list of decisions a room has taken: approvals, refusals, mode changes and reverts, with who and when, extracted from the timeline.

Live preview

The room's running app, shown beside the timeline to every member while it is being built.

Publish

Putting a static snapshot of the room's app at a public address. Owner only, explicit, and recorded in the timeline. Apps that need a server are previewable but not yet publishable.

Watch link

A view-only link to a room's timeline and group chat that needs no account. Any member can turn it off at any time, which ends every viewer's access at once.

Multi-agent mode

A room setting where each member drives their own agent on their own branch, merged into the room's primary branch when ready, instead of sharing one agent.

Egress approval

What happens when a turn tries to reach a domain outside the allowlist: the request is held for up to two minutes while a card names the domain, and any member allows or denies it. Nothing is approved automatically.

Credits

Poly's unit of billing for agent usage, charged to the person who ran the turn and shown on its receipt. Every account starts with free credits; watching, chatting and approving never cost any.