The Multiplayer AI Manifesto

The first generation of AI inherited the interaction model of legacy client-server software: single person, one interface, single session. A user opens a window, asks a question, gives the model some context, receives an answer, and closes the window. The model is very smart and will reason across thousands of tokens, use tools, write code, analyze data, and execute increasingly complex tasks. But the unit of interaction remains fundamentally individual.

However, most real work in companies is not performed by isolated individuals. It is performed by groups of people who share information, divide responsibilities, inherit decisions, challenge one another, and coordinate over time. Organizations function through networks of collaboration.

However, it seems like we have built increasingly sophisticated agents and placed them inside increasingly sophisticated silos. We need Multiplayer AI.

What is Multiplayer AI?

Multiplayer AI is a model of computing in which people and agents work together in shared, persistent environments toward common objectives.

A multiplayer system is not simply a chatbot with multiple users. It changes the fundamental interaction model from one human interacting with one machine → groups of humans and agents coordinating with one another.

• A multiplayer system may contain several people working with the same agent. 

• It may contain several agents collaborating on different parts of a problem.

• It may contain agents operating on behalf of different teams, with different permissions, responsibilities, tools, and expertise.

• And increasingly, it will contain all three simultaneously.

The resulting system is less like a chatbot and more like a networked organization.

Software became collaborative over time but AI chatbots took us backward.

The history of modern software is : Documents became collaborative documents, Source code became shared repositories, Projects became shared workspaces, Communication became persistent channels rather than private messages, Design became multiplayer and so on

The most important productivity systems of the last twenty years succeeded because they made work shared, inspectable, persistent, and composable.

Then generative AI arrived. And much of the computing returned to the private terminal.

Single user. Single conversation thread.One ephemeral context window. One agent that knows what you told it, but not necessarily what your colleague told another instance of the same agent five minutes earlier. 

Context tax should be repealed

Before useful work can begin, humans must reconstruct the relevant state of the world. For e.g when discussing a customer issue →here is what happened last week, the constraints, what the legal team said, what we already tried, what the previous agent recommended, what changed since then.

The issue is not the model's ability to think but rather shared context. Organizations already possess enormous amounts of context. It exists in conversations, applications, documents, databases, decisions, exceptions, workflows, relationships, and accumulated experience.

Private AI conversations fragment it further.

A person asks an agent to analyze a customer. The agent produces an excellent answer. The answer is pasted into Slack. Another employee asks another agent to continue the work. The original context is lost. Humans become the synchronization protocol between intelligent systems. This is a poor use of humans.

People should not be routers

A surprising amount of knowledge work consists of moving context between systems and people. This happens because our systems do not share state effectively. Agents make this problem 10X worse.

If an agent is capable of finding information, interpreting it, and communicating with other systems, there is little reason a person should spend their time manually routing information between them.

Humans should provide judgment, direction, creativity, accountability, and disagreement. They should not function as middleware.

Shared context should amplify intelligence

The capabilities of an agent are usually discussed as properties of the model.

How well can it reason? Which benchmarks can it solve?

But in an organization, intelligence is not solely a property of an individual reasoner. It is also a property of the environment in which reasoning occurs. An agent operating inside a persistent shared workspace can know what the team has already discussed.

• It can observe previous decisions

• inherit the state of an unfinished task, 

• incorporate corrections made by another person, 

• understand which assumptions have already been rejected, 

• recognize when two teams are solving the same problem independently, 

• ask another agent for specialized assistance.

It can act without requiring the organization to reconstruct its world from scratch each time.

The shift from single-player AI to multiplayer AI is an architectural change.

Shared state compounds intelligence.

Agents will collaborate with agents

The multiplayer model extends beyond human collaboration.As agents become more capable, no single agent will perform every task. Organizations will deploy agents with different capabilities, permissions, tools, models, and responsibilities.

• Research Agents may gather evidence.

• Analytical Agents may interrogate structured data.

• Customer Agents may understand account history.

• Financial Agents may evaluate commercial implications.

• The Governance Agent may inspect the proposed action before execution.

The important problem will no longer be simply How effectively can this system of agents coordinate?

This introduces classical problems from distributed systems into AI: coordination, state, identity, permissions, failure recovery, observability, provenance, synchronization, delegation, and conflict resolution.

Work should have memory

Most AI sessions today are treated as disposable interactions. Work should not be disposable.

A serious system must preserve the lineage of a task, This history is not merely an archive.It is computational memory. A task completed today should make the system better prepared for the same task tomorrow. A successful workflow should be reusable.

In multiplayer AI, work accumulates.

Collaboration requires security

Shared intelligence cannot mean unrestricted intelligence. The more deeply agents participate in organizational work, the more seriously we must treat identity, authority, and access.

An agent should know not only what information exists, but what information it is permitted to use for a particular person and task.

The same is true of actions. Which systems may it access? What data may it reveal? Which actions may it perform? They are part of the semantics of the work itself.

Multiplayer AI therefore requires security models that understand humans, agents, teams, resources, and delegated authority as first-class entities.

The principles of Multiplayer AI

We believe the next generation of intelligent software should be built around several principles.

1. Never copy and paste context  Agents should participate in the systems and conversations where work already occurs. Context should travel with the work rather than being repeatedly reconstructed by humans.

2. Work with the door open. AI-assisted work should be visible to the people who need to understand, challenge, continue, or learn from it. Private reasoning has its place. Organizational work should not disappear into private transcripts by default.

3. Nothing important should start from zero. Sessions should persist. Decisions should have history. Agents should be able to resume work with the state, evidence, and constraints that already exist.

4. Feedback should compound. A correction should improve more than one conversation. Repeated instructions should become reusable knowledge, skills, policies, or workflows.

5. People are not routers. Agents should retrieve context, coordinate routine work, communicate state changes, and hand tasks to appropriate humans or agents.

6. Agents should be able to ask other agents for help. Complex work should be decomposable across specialized agents without requiring a person to orchestrate every intermediate step.

7. Authority must remain explicit.An agent's ability to know or act must derive from identifiable people, policies, and permissions. Delegation should be visible and revocable.

8. Intelligence should be inspectable. Teams must be able to understand what an agent did, what information it used, what decisions were made, and where uncertainty remains.

9. Organizational knowledge should outlive individual sessions. The useful residue of work should become part of the organization's collective memory rather than remaining trapped in personal chat histories.

10. Humans remain participants, not spectators.Multiplayer AI is not about removing people from work. It is about changing where human effort is applied: away from mechanical coordination and toward judgment, creativity, negotiation, direction, and responsibility.

From personal chatbot to team intelligence

Single-player AI is extraordinarily useful. It makes an individual faster. The larger opportunity is therefore not merely to give every employee an increasingly capable private assistant. It is to create environments in which the intelligence of people and machines can accumulate, coordinate, and improve together.

That requires shared context. Shared work.  Shared memory. Shared agents. Explicit permissions. Persistent state. And the ability for humans and machines to enter and leave a task without destroying its continuity.

The interface of the furure will be people and agents operating together. That is Multiplayer AI.