Multiplayer Agents

What people mean by “multiplayer agent”

In single-player mode, each person has a private chat with an AI. Context lives in their thread. When a teammate needs the same work, they re-explain from zero — or paste a screenshot of a half-finished answer.

Multiplayer agent means: multiple humans work with the same agent identity (or the same role agent) in a shared surface, and the agent keeps memory and context that outlive any one person’s session.

• Alex asks the content agent for three hooks → agent drafts in a team room
• Sam replies “use the stronger second one, and match last week’s voice rulings”
• Agent already knows the voice file, last week’s rulings, and the channel history
• Nobody re-pasted the brand guide; the third person tomorrow can still pick up the thread

Slack marketed Claude Tag as “AI that’s multiplayer, where work already happens.” Eric Siu’s Hermes ladder puts the same idea at Level 6: multiplayer + connected data — see Hermes as an operating layer. Company Brain is the org-scale version: one ambient intelligence the team talks to in the tools they already use.

Single-player vs multiplayer (quick contrast)

The architecture (four layers)

Multiplayer is not “put a bot in Slack and hope.” It is four layers wired together:

Humans A, B, C
      │  @mention / thread
      ▼
[ Shared room ]  ← visible to team
      │
      ▼
[ Role agent ]  ← identity + tools + schedule
      │
      ├── loads slice of [ Knowledge vault / HEOS ]
      ├── writes artifacts + optional rulings
      └── gated actions wait for human approve

What “memory and context” actually means

Friends often say “keep memory” as one blob. In practice there are three kinds — mix them on purpose:

Rule of thumb: if two people should still know it next month without reading 400 messages, it belongs in the vault (or a skill with a “done” definition) — not only in the multiplayer thread. When you correct the agent in front of the team, add one line to rulings so the next human inherits the fix (fix once, not forever).

How multiplayer feels day-to-day

1. Someone starts work in the open. “@ContentAgent draft a CBS sponsor email using voice.md + this offer snippet.”
2. The agent answers in the room so others can see drafts, not only the requester.
3. Teammates steer with short corrections (“shorter,” “no hype words,” “use the Q3 pricing”). Those corrections should become vault/rulings when durable.
4. Artifacts get pinned or filed (checklist, HTML one-pager, approved copy) so the room isn’t the only archive.
5. Next person continues without re-uploading the brand bible — agent loads the same slice + recent room context.
6. Crons report into the room (“pipeline quiet deals,” “standup blockers”) so multiplayer isn’t only human-initiated.

That “Chris asks for open leads → I see the answer → I ask a follow-up” loop is exactly how Eric describes multiplayer Hermes in Slack. We want the same loop in GFAVIP.

Patterns that work

Anti-patterns

• One mega-agent in #general for every question — noise, wrong tools, no slice discipline.
• Memory only in chat history — switch apps or hit retention limits and the “team brain” evaporates.
• Everyone’s personal ChatGPT + occasional paste into Slack — looks collaborative; isn’t multiplayer.
• Send/spend tools with no human gate — multiplayer amplifies mistakes (everyone saw it ship).
• Identity mixed with memory — a bad “note to self” rewrites who the agent is for the whole team. Keep identity files stable (KB / Hermes).
• Member-facing VIP agents sharing internal ops rooms — different trust boundary; see GFAVIP AI Agents.

How this maps to our stack

Minimal setup checklist (teach the team this)

□ Pick one role agent (e.g. Content) — not “the company bot”
□ Create one GFAVIP room; invite humans who own that work + the agent account
□ Pin vault slice: offer / voice / playbooks / rulings relevant to that role
□ Tools: read + draft first; no external send until a gate exists
□ Rule in system prompt: push back, rate ideas, improve the system
□ After durable corrections: one line in rulings.md (or decision-log)
□ Publish artifacts out of the room into HEOS / docs when “done”
□ Name a human DRI for the agent (quality, cost, incidents)
□ Optional: morning cron posts into the same room
□ After 2 weeks: add a second role room — only if the first is used without re-briefing

FAQ

Is multiplayer the same as multi-agent?

No. Multiplayer = many humans ↔ one (or few) shared agents. Multi-agent = many agents coordinating (research + writer + reviewer). You usually want both: multiplayer rooms and specialist agents that hand off.

Should every person get their own multiplayer agent?

No for company work. Role agents are shared. Optional personal chief-of-staff stays private (allocation decision). Multiplayer is for work that benefits from visibility.

Won’t the room get noisy?

Use threads for long runs, pin artifacts, keep one job per room, and kill stale crons. Noise is a room-design problem, not a reason to go back to private silos.

Where does context lock-in show up?

If the only place decisions live is Slack/Claude’s agent memory, you rent your brain. Multiplayer surface can be vendor; source of truth should be your vault/OS (full write-up).

Related on this site

References & further reading

• Eric Siu on Hermes multiplayer / company brain ladder — our Hermes guide · source thread
• Eric Siu — Company Brain talk notes
• Slack / Anthropic multiplayer framing — Claude Tag analysis
• GFAVIP — gfavip.com/download · Webchat guide