The Multiplayer AI Manifesto

What we think AI at work should look like. Five principles for multiplayer AI: never copy-and-paste, work with the door open, continuously improve, people are not routers, nothing starts from scratch.

Why we need this manifesto

Reason

We went from collaborative work to siloed AI chats

Collaborative features are a given in today’s apps. We went from email to Slack. From Word to Google Docs. But recently, we have drifted apart into individual AI chats.

Each of those chats is a closed room: whatever gets figured out in there stays with just one teammate by default.

Reason

Siloed chats come with a context tax

Today, Priya might work with Claude on a report, then send it to Marcus via Slack, only for him to put the report into ChatGPT and email the answer back.

With each handoff, the company pays a context tax that they wouldn’t have to pay if teammates could collaborate in the same agent session.

Reason

AI agents require org-wide security

AI has gone from answering questions to taking actions on your systems.

An AI agent running on an employee’s machine might inadvertently move sensitive data from one system it’s connected to into another — or worse, onto the public Internet.

Reason

Multiplayer AI improves outcomes

When employees of a large enterprise were allowed to collaborate both with each other and with AI, they achieved the highest proportion of top-quality outcomes in a Harvard Business School study.

And companies like Shopify found that allowing all employees to chat with AI agents in public was a boon to productivity. Within a month, 1 in 8 pull requests was already co-authored by their AI agent River.

Common misconceptions

Misconception

Isn’t my AI enterprise plan already multiplayer?

No, ChatGPT and Claude team/enterprise plans only let you share a few settings: prompts, skills, and integrations. They do not allow true multiplayer in the sense that Slack, Google Docs, Figma, and other modern tools do.

There is no chatting with a teammate’s agent. There is no learning from each other. Context is trapped in individual sessions, unindexed and short-lived.

Misconception

Can’t I just add an agent to Slack?

The simplest solution that allows your team to work with agents together is adding them to Slack. And indeed, Claude Tag and other solutions are great starts.

But a lot of work happens off Slack: in emails, Google Docs, GitHub, Salesforce, ServiceNow, etc. Truly multiplayer AI must enable reaching the same agent session from every work surface and every device.

We posit the following five principles for a true Multiplayer AI experience. This is a North Star to build toward. No vendor or internal platform that we are aware of currently satisfies all of them.

1. Never copy-and-paste
2. Work with the door open
3. Continuously improve
4. People are not routers
5. Nothing starts from scratch

Principle

Never copy-and-paste

• Agents must live next to the work, instead of in their own chat or in Slack.
• Every person who is involved in the work must be able to chat directly with the same agent.
• Agents must be able to use every tool that a human on the team uses to do the work.

The same work as the context tax relay above, in one session.

Principle

Work with the door open

• When chat sessions stay private, best practices spread slowly. When chat sessions are shared, everyone gets better quickly.
• Tobi Lütke, Shopify CEO, called this “learning on the shop floor”. Shopify’s internal AI system (River) is only accessible via public Slack channels.

…if you have the door to your office closed, you get more work done today and tomorrow, and you are more productive than most. But 10 years later somehow you don’t quite know what problems are worth working on; all the hard work you do is sort of tangential in importance.

Principle

Continuously improve

• When a teammate writes a great prompt and the agent gets it right on the first try, other teammates should learn how to write good prompts.
• When an agent needs correcting, a reusable skill should be created automatically, a saved recipe for that task that future agents use.
• When the same workflow is performed repeatedly, a benchmark should be automatically created and its score tracked.

Principle

People are not routers

• A human must never be asked a question that an agent already has the answer to.
• Chasing project updates and relaying answers is work for an agent, not a human.
• Humans must focus only on the highest value activities. The bureaucracy must be digitized.

Principle

Nothing starts from scratch

• When agent sessions live in the cloud, you can pick up every project exactly where you left off, with no context lost.
• Every artifact, whether a design doc, financial plan, or pull request, must have an agent session that can be resumed even months later.
• A new person joining a multiplayer AI team is productive on day 1.

Important considerations

Whether you build or buy, multiplayer AI is tricky to get right.

Consideration

Agents don’t belong on laptops

• Laptop agents can’t be accessed by the rest of the team.
• Laptop agents stop working when the lid gets closed.
• Laptop agents can gain access to everything that an employee has access to.

Consideration

Firewall every agent

• An agent tricked by a malicious email or document can send sensitive data to an outsider. Security people call this prompt injection.
• Every agent must be limited to the services you approve. That is only practical in locked-down cloud environments.

Consideration

Privacy pitfalls

• Agents are at their most useful when they have access to sensitive data such as emails, Slack messages, financial records, etc.
• When adding another teammate to a session that has exposure to sensitive data, it is important to handle permissions correctly.
• For example, the agent should only have access to integrations that every member of a shared session has access to.

Consideration

Stay provider-agnostic

• Some teammates prefer Claude, some prefer ChatGPT, and your CFO prefers open source.
• The most powerful model changes week by week, and most tasks do not require the most powerful model in the first place.
• When you bet your company on one model provider, you align yourself with their incentive structure, which is to bill you for as much usage as possible.

Consideration

Governance is non-negotiable

• Governance means being able to answer four questions about any session a year later: which agent ran, who asked it to, what data it touched, and what it changed. Regulated industries and anyone subject to GDPR cannot skip this.
• Auditors need a “black box” for agents, just as they have for every system of record: enough to check what happened without redoing the work by hand.

Session #4128 · audit recorda year later

Data accessedAcme contract, 3 decks, 4 ticketsChanges madeNone. A draft was saved.

Consideration

Build, buy, or open source

• There are compelling reasons for each one of building internally, running an open-source solution, or using a hosted platform.
• Most organizations find that while a basic version can be built quickly, getting to a fully-featured, seamless experience is a significant effort that detracts from their main goals.
• Whichever way you go, one thing is not up for negotiation: you must own your data — the learnings, skills, and evals your team creates by using the platform.

Compare solutions

No platform satisfies all of these principles at this time. But some are closer than others. We are building a detailed comparison, featuring leading platforms such as Buzz, Claude Tag, Superconductor, Viktor, and more.

Questions? Comments? Please reach out.