The next enterprise AI problem is not getting more people into the conversation. It is giving people and agents a shared reality in which consequential decisions can be made.
The recently published Multiplayer AI Manifesto identifies something important.
As AI has become more capable, work has become strangely less collaborative.
We spent decades moving enterprise work out of individual files and inboxes and into shared environments. Documents became collaborative. Software development became collaborative. Communication became collaborative.
Then AI arrived – and much of that work went back behind closed doors.
One employee works something out with an agent. Another receives the result, opens another AI session, reconstructs the context, asks another question, and sends the result somewhere else.
The Multiplayer AI Manifesto calls the resulting friction the “context tax.” That is a useful description. But it may also point toward a deeper problem.
Because once agents begin participating in consequential enterprise decisions, the question becomes: ‘What shared reality are those people and agents actually operating against?’, not just ‘How do we give everyone access to the same conversation?’
That distinction matters.
The Context Problem Is Real
The five principles proposed by the Multiplayer AI Manifesto are difficult to argue with.
- Agents should live close to the work.
- Knowledge should not disappear inside private AI sessions.
- Good practices should accumulate rather than continually being rediscovered.
- People should not spend their time acting as routers between systems and other people.
- And work should not repeatedly begin from zero.
Perhaps most importantly, the manifesto recognizes that governance cannot be added later. If an agent participated in enterprise work, the organization should eventually be able to determine who invoked it, what information it accessed, and what it changed.
These are meaningful requirements for enterprise AI.
They also expose the limitations of the dominant AI interaction model.
A collection of highly capable agents operating inside isolated sessions is still a collection of isolated actors.
Making those sessions multiplayer improves the situation considerably.
But collaboration is not yet orchestration.
From the Context Tax to the Coordination Tax
The Book makes a related argument from a different starting point.
The central constraint emerging inside the AI-enabled enterprise is increasingly not intelligence.
It is coordination.
Enterprises already possess extraordinary amounts of intelligence: forecasts, optimization engines, analytics, automation, AI assistants, agents, domain applications, and human expertise.
Adding another intelligent component can improve one part of the organization without necessarily improving the behavior of the organization as a whole.
Local intelligence does not automatically produce system intelligence.
That is why the context tax described by Multiplayer AI is so interesting.
It may be the visible symptom of a larger coordination tax.
Consider a deceptively simple question:
Can the pricing page ship Thursday?
A shared agent session might know the specification, inspect the pull request, discover a discrepancy, make the correction, notify marketing, and preserve the entire conversation for everyone involved.
That is substantially better than passing fragments of context between private AI sessions.
But now change the question:
Should we accept a $2 million customer commitment under these terms?
The nature of the problem changes.
The answer may depend on inventory availability, manufacturing capacity, logistics exposure, contractual obligations, service levels, margin thresholds, existing commitments, customer priority, current disruptions, corporate policy, and several plausible versions of what happens next.
Some of those facts may be uncertain. Some may conflict. Some may change while the decision is being made.
And several different actors – human and machine – may have legitimate but competing objectives.
At that point, the enterprise does not merely need shared conversational context.
It needs shared operational state.
A Session Remembers the Conversation. A World Model Remembers the Enterprise.
This is where the distinction becomes architectural.
Multiplayer AI proposes persistent sessions connecting people, agents, tools, and artifacts.
An orchestration-native enterprise requires something complementary but deeper: a persistent, governed representation of the operational world those participants are attempting to change.
That world must represent more than documents and messages.
It must represent things such as:
state → evidence → intent → constraints → alternatives → simulation → policy → decision → commitment → action → outcome
The important difference is not semantic.
A conversation can tell us what people discussed.
A world model must tell us what the enterprise currently believes to be true.
A conversation can preserve an agent recommendation.
A decision runtime must preserve the evidence on which that recommendation depended, the alternatives considered, the policy boundaries applied, the authority under which action was permitted, and the resulting change in operational state.
A conversation can be resumed.
An operational world must be reconstructed and replayed.
That is a much stronger requirement.
“People Are Not Routers” Raises a Bigger Question
One of the strongest principles in the Multiplayer AI Manifesto is also one of the most consequential:
People are not routers.
Agreed. Humans should not spend their working lives discovering information in one system and carrying it into another.
Nor should highly paid people exist primarily to chase status updates, reconcile conflicting reports, or manually propagate decisions through an organization.
But removing the human router raises an architectural question:
What becomes the router?
The obvious answer is the agent.
Give agents sufficient context, connect them to enough systems, and allow them to move information and initiate work.
That solves part of the problem.
But it risks reproducing the same architecture with a more capable intermediary.
Instead of humans carrying information between fragmented systems, agents carry information between fragmented systems.
The router became faster. The enterprise did not necessarily become more coherent.
The orchestration-native alternative is different. Nothing should merely “route” the enterprise.
The system should continuously maintain an explicit model of operational reality against which humans, agents, policies, simulations, and systems of record can coordinate.
The organizing primitive moves from workflow to state.
And increasingly, from conversation to state as well.
The Difference Becomes Most Visible at the Moment of Commitment
AI is very good at producing possibilities.
Enterprise operations ultimately require commitments.
- A forecast is not a commitment.
- A recommendation is not a commitment.
- A simulation is not a commitment.
- An agent proposal is not a commitment.
- Even an agreement reached in a shared conversation is not necessarily an enterprise commitment.
A commitment changes what the organization is now obligated – or authorized – to do.
That transition requires something more rigorous than intelligence.
The system must know:
- What is currently true?
- What are we trying to accomplish?
- What constraints apply?
- What alternatives exist?
- What happens if we choose each one?
- Which policies govern the decision?
- Who – or what – has authority to commit?
- What changed as a consequence?
- And can we reconstruct that decision later?
This is the territory of orchestration (the i5 definition, anyway).
Governance Must Move from the Agent to the Decision
The Multiplayer AI Manifesto is particularly aligned with i5 on another point: governance is non-negotiable.
Its proposed “black box” for agents asks sensible questions:
- Which agent ran?
- Who requested it?
- What data did it access?
- What did it change?
Those questions will become table stakes.
But consequential enterprise decisions require several more:
- What was the state of the world when the decision was made?
- What evidence supported that state?
- What alternatives were considered?
- What simulations were performed?
- Which policies constrained the decision?
- What authority permitted the commitment?
- What actually happened afterward?
This moves governance beyond an agent audit trail. It creates a durable decision and provenance record.
That distinction becomes increasingly important as enterprises move from AI that recommends actions toward AI that participates in making and executing them.
Trust cannot depend on reconstructing intent from chat transcripts after something goes wrong. Trust has to become a property of the runtime.
Multiplayer May Be the Right Interface. World State May Be the Missing Substrate.
None of this makes the Multiplayer AI idea wrong. Quite the opposite. It suggests something potentially powerful about where enterprise AI architecture is heading.
Multiplayer AI may describe an important interaction layer for the emerging enterprise:
- People and agents working together.
- Shared sessions.
- Persistent context.
- Multiple work surfaces.
- Accumulating organizational knowledge.
- Provider-independent intelligence.
INDUSTRY 5 is concerned with the layer beneath that interaction:
- Shared world state.
- Evidence.
- Intent.
- Constraints.
- Simulation.
- Policy.
- Authority.
- Commitments.
- Actions.
- Outcomes.
- Replay.
Put differently: Multiplayer AI addresses where humans and agents meet. Orchestration addresses the reality against which they decide and act.
The two ideas are not competitors.
They may be different layers of the same emerging architecture.
Intelligence Is Becoming Abundant. Coordination Is Not.
There is a broader pattern here.
The first generation of enterprise AI focused on intelligence. Give the employee a smarter assistant.
Then came agents. Give the assistant tools and let it act.
Multiplayer AI recognizes the next problem.
Give those agents and humans shared context so intelligence stops disappearing into isolated sessions.
But there is another transition beyond it.
Give the enterprise a persistent, governed model of itself so that all of those intelligent actors can coordinate decisions against the same operational reality.
The progression looks something like this:
Individual intelligence
↓
Agentic action
↓
Shared context
↓
Shared operational state
↓
Governed orchestration
That final transition is the wager behind i5.
The enterprise of the future will certainly contain agents.
It will probably contain many of them.
They will collaborate with people, invoke tools, use different models, participate through many interfaces, and increasingly operate continuously rather than waiting for prompts.
But agents themselves are not the operating model. They are actors within it.
In some respects, this points toward what is increasingly described as a neurosymbolic architecture: combining the probabilistic intelligence of AI with explicit, deterministic representations of state, relationships, constraints, policy, and authority.
The distinction matters because agents can reason about the operational world without being asked to invent that world each time they reason.
The harder problem is constructing the environment in which those actors can understand what is true, reason about what could happen, negotiate competing objectives, operate within explicit authority, make commitments, observe consequences, and learn without sacrificing governance.
That requires more than multiplayer intelligence.
It requires a shared, governed world.
And that may prove to be the real architecture of the orchestrated enterprise.
Further Context
The Multiplayer AI Manifesto is worth reading in full, particularly its principles around persistent context, human routing, provider independence, and governance.
For the broader argument about why coordination – not intelligence – is becoming the limiting constraint in enterprise AI, see The Orchestrated Enterprise.
And for the implementation direction behind i5 – explicit world state, simulation, governed commitments, append-only events, and deterministic replay – see i5 World-Runtime.