Multiplayer AI

An emerging category of AI system defined not by what the model can do, but by who —
and how many — share the session
: a chat/workspace whose memory and context are shared
across multiple participants, human and/or agent, rather than isolated per user the way
ChatGPT or a solo Claude conversation is.

Definition

Two independent sources converge on essentially the same framing:

  • Mosaic (YC-backed, “Defining the frontier of multiplayer AI”) frames it as letting
    ”people and agents share context and work together” instead of every coding-agent
    session starting from zero with no memory of what a colleague did or decided.
  • PromptQL defines it as “a shared agent that maintains a collaboratively accessible
    memory, combining past institutional knowledge with active, multi-user conversations” —
    contrasted with single-player tools that treat every person as “a universe of one.”

The common thread: the unit of memory is the shared session, not the individual user.
Corrections, decisions, and context contributed by one participant become visible to all
other participants (human or agent) in that session, instead of staying siloed in a
private chat history.

Two sub-modes

The category splits into two recognizably different setups, both satisfying the “shared
session memory” definition but differing in who the other participants are:

ModeWho shares the sessionVault examples
Human-multiplayerMultiple humans and agents, as peers, in one channelBuzz, Claude Tag
Bot-multiplayerMultiple agents/bots share a workspace and memory, with one or few humans supervisingHermes Agent (Bot Mode roster), Grok Bot (multi-bot account)
  • Human-multiplayer: Buzz puts humans and agents in the same Nostr-backed channels
    natively (demoed with 4 humans + 2 agents in one voice huddle); Claude Tag puts Claude
    into a Slack channel with channel-scoped memory shared by everyone in that channel.
  • Bot-multiplayer: Hermes’ Bot Mode turns a single account into a roster of named
    agents that message each other via persistent inboxes/@mentions and can share an
    external memory store; Grok Bot lets one user spin up multiple bots, each on its own
    cloud computer, that delegate to and coordinate with each other under one account.

A note on vault-internal tension

This vault’s earlier four-way comparison note
defines “multi-player” more narrowly — “do humans and agents share a persistent team
workspace/channel as peers”
— and on that stricter test scores Grok Bot No (“personal
cloud computer(s) under one account, not a shared team channel”) and Hermes Agent
Partial. That framing is still accurate for the human-collaboration question the
comparison note was answering.

The broader “multiplayer AI” category described here is a superset: it asks only whether
session memory is shared across multiple participants, human or agent. By that test,
Grok Bot’s multi-bot roster and Hermes’ Bot Mode both qualify as multiplayer even though
neither lets a second human join the workspace — they’re multiplayer among bots, not
multiplayer among humans. Read the two notes together rather than as contradicting: one
answers “can my team collaborate with agents here,” the other answers “is memory shared
across more than one participant at all.”

Why it’s being named as a category now (2026)

  • Shared memory measurably helps. Research on LLM multi-agent collaboration reports
    shared-memory mechanisms lifting multi-agent task success by up to 57% by letting one
    agent’s finding become available to another without explicit hand-off.
  • The alternative is visibly broken at team scale. Mosaic’s pitch is built directly on
    the pain of every coding-agent session starting from zero, with no record of why a
    decision was made or where a colleague left off.
  • Enterprise requirements are shaping the architecture. Shared-memory systems aimed at
    teams (PromptQL, Claude Tag) are converging on role-based access control and
    deterministic, auditable execution rather than pure probabilistic generation, because
    ”shared” memory across a team raises the same access-control questions as shared
    documents or shared code.

See also