— AI Can Share Knowledge, but It Cannot Hold a Meeting —
In the previous article, I wrote about the possibility that AI has added a new layer to Knowledge Management.
If Knowledge Management deals with how knowledge is created, accumulated, shared, and applied within an organization, then the participation of AI introduces a new set of design questions:
- Who refers to the knowledge?
- Which knowledge is selected?
- To whom is it delivered?
- Into what form is it transformed?
I proposed thinking of this behavior as Knowledge Routing.
However, while working on a single project with multiple AI systems, I encountered a phenomenon that Knowledge Routing alone does not seem sufficient to explain.
Even when Knowledge is shared, the state of collaborative work does not return.
Building a Game with Multiple AIs
I recently developed a game based on Reversi in collaboration with multiple generative AI systems.
Each AI had a different role.
One AI worked on planning and specifications, another handled implementation, and another was responsible for moderation and information organization.
In a human development team, this would be similar to having a planner, a programmer, and a facilitator working together.
This kind of development is already quite possible with current AI systems.
They can produce specifications, write code, run tests, fix problems, and ultimately complete a working piece of software.
However, when I wanted to change the direction of the project during development, a problem emerged.
All participants needed to discuss the change together.
With Humans, You Can Simply Call a Meeting
In a human team, this would not be difficult.
“Let us get together for a few minutes.”
You would simply open a video conference.
Each participant would bring whatever they had been thinking about immediately beforehand.
What they considered important.
Where they believed the problem was.
Which proposal they considered promising.
What they had not yet decided and were intentionally leaving unresolved.
They would enter the discussion while retaining those states, listen to the other participants, and update their own thinking.
When the meeting ended, each person would return to their work with an updated understanding.
Humans do not usually consider this mechanism particularly remarkable.
It happens naturally when a meeting is held.
AI Cannot Join the Same Meeting
Current AI systems cannot do the same thing.
Each AI exists inside a separate conversational environment.
A conversation with ChatGPT contains the context accumulated with ChatGPT.
A conversation with Claude contains a different context accumulated with Claude.
A conversation with Gemini has yet another context formed within that environment.
Even when all of them are discussing the same project, their conversation histories and the current states formed from those histories are not identical.
Of course, documents can be shared.
Specifications can be passed between systems.
A Knowledge Base can be prepared.
Using MCP and similar mechanisms, they can also be given access to the necessary information and tools.
But even then, it is not the same.
Passing Knowledge does not restore the state in which the participants had been thinking together.
This was the problem I encountered while actually developing software with multiple AI systems.
What Is Not Stored in a Knowledge Base
What can be stored in a Knowledge Base is, fundamentally, knowledge.
Confirmed specifications.
Design decisions.
Problems that have occurred in the past.
The conclusions of discussions.
These can be stored as documents or data and passed to another AI.
However, during collaborative work, there are things that have not yet become documents.
For example:
- Which problem currently holds the participant’s attention
- Which hypothesis the participant currently considers most plausible
- Which explanation feels unsatisfactory
- What is being intentionally left unresolved
- In which direction the participant is about to continue thinking
These are not simply pieces of knowledge.
Nor are they equivalent to the complete conversation history.
Even if another agent reads the same conversation history, the same focus, concern, or line of inquiry will not necessarily be reproduced.
There appears to be another kind of state: a state describing how knowledge is currently being noticed, connected, and used.
Could There Be an Activation State?
To explain this phenomenon, I have begun to wonder whether some form of activation state might exist.
One might also call it an excited state.
However, no clear definition of such a state currently exists.
What, exactly, is being activated?
What kind of information does the state contain?
To what extent can it be stored, transferred, or restored?
Can it even be treated as an independent state?
At present, I do not know.
The point here is not that I have defined what an activation state is.
The question is:
Could there be some kind of state that cannot be fully explained by Knowledge or Context alone?
Even when multiple AI systems are given the same Knowledge Base, the collaborative thinking they had been engaged in immediately beforehand is not restored.
By contrast, humans can, at least in practical terms, bring that state back by gathering for a meeting.
This difference does not appear to be merely a difference in the amount of information available.
AI Has Made the Role of Human Meetings More Visible
Meetings are often described as mechanisms for sharing information.
However, if the purpose were only to share information, documents and email would often be sufficient.
Even so, when an important situation arises, humans gather and talk.
Perhaps this is because they are not merely exchanging information.
They may also be reactivating one another’s thinking, allowing those thoughts to interact, and synchronizing their current understanding.
In organizations consisting only of humans, this function may have been so natural that it was rarely recognized as an independent problem.
When working with AI, however, its absence becomes clearly visible.
AI makes this problem more explicit than humans do.
With humans, one can simply call a meeting.
With AI, there is currently no way to bring multiple systems into the same meeting while preserving the states accumulated in their respective conversational environments.
Through collaboration with AI, we may be seeing more clearly what human meetings were actually doing all along.
Current Orchestration Is Not Enough
Today, APIs and MCP-based orchestration are often considered as ways of coordinating multiple AI systems.
These mechanisms can allow one AI to call another, or give multiple AIs access to a shared Knowledge Base and a common set of tools.
Tasks can be divided, assigned to different systems, and their results can be integrated.
However, in many such architectures, each AI begins its work after being given the necessary documents for that particular task.
The context accumulated in each AI’s original conversational environment, and the state of thought immediately preceding the call, are usually lost.
This may be similar to giving minutes and specifications to a worker who has never participated in the previous discussion, and then asking that person to continue the work.
The documents exist.
The necessary knowledge is written down.
But this is not the same as continuing with the participant who had been discussing the issue with you until yesterday.
What I felt I needed was not a new API session calling an AI from scratch.
I needed each AI to join a meeting from its current conversational environment while preserving its existing state.
Then, after the meeting, it should return to its original conversation and task while carrying the results of that discussion.
Why Coordination May Now Have Real Potential
There have already been many attempts to coordinate multiple systems and multiple AI agents.
They have not always produced good results.
My view is that one reason may not have been a fundamental flaw in the idea of coordination itself, but the insufficient capabilities of the individual functions being coordinated.
In earlier systems, speech recognition, image recognition, natural language processing, and reasoning were all highly sensitive and narrow functions.
They might work under tightly constrained conditions, but once an assumption failed, they could no longer produce a correct result.
As a result, integrating them with other functions required the entire system to be designed around the limitations of the most sensitive component.
In some cases, it may be more accurate to say that the other components were supporting and compensating for the most unstable function, rather than truly coordinating with it.
Current AI systems, however, are beginning to acquire the ability to interpret imperfect output and ambiguous expressions, and to connect them to other systems.
LLMs can be used not only to replace specialized functions, but also as adapters that sit between different functions and absorb ambiguity.
This means that individual specialist systems no longer need to be universal.
Even a relatively small specialist AI may be able to participate as a member of a team, provided it can be meaningfully connected to other AI systems.
A Future Beyond a Single Giant AI
If effective coordination becomes possible, the future may not belong only to a single giant AI.
It may also include architectures in which multiple, comparatively small specialist AIs participate together.
For example:
- An AI responsible for judgments within a specialist domain
- An AI responsible for implementation
- An AI responsible for information retrieval and organization
- An AI responsible for moderating the discussion
- An AI responsible for explaining the result to humans
Each could participate in the same project with a different role.
In such a system, the most important question would not be which AI is individually the most intelligent.
The important question would be:
How can different forms of intelligence preserve their states and influence one another while working together?
Sharing a Knowledge Base would not be enough.
Delivering the necessary knowledge through Knowledge Routing would not be enough either.
Each participant has a current focus, a set of concerns, and a collection of hypotheses.
A mechanism may be required that allows them to participate in collaborative thinking while preserving those activation states.
The Question That Appeared After Knowledge Routing
In the previous article, I proposed the idea of Knowledge Routing.
Who does knowledge pass through?
In what form?
To whom is it delivered?
I believed that if AI was going to participate in Knowledge Management, a layer for designing those routes would become necessary.
After actually developing software with multiple AI systems, another question appeared beyond that layer.
What happens inside an agent after knowledge has been delivered?
Why can the state of collaborative thought not be restored from a Knowledge Base alone?
What are humans reactivating when they hold a meeting?
What would need to be stored and connected in order for AI systems to join a meeting while preserving their individual contexts?
Could some form of activation state be involved?
I do not yet know the answer.
I do not even know whether “activation state” is the right term.
It is also unclear whether this can be defined as an independent concept.
However, when actually working with multiple AI systems, there appears to be something that cannot be fully explained by Knowledge or Context alone.
This is not only a question about AI.
It is also a question about the nature of human coordination.
From Knowledge Management to Knowledge Routing.
And beyond that, there may be another field concerned with the activation states of knowledge-bearing agents and the mechanisms by which those states are synchronized.
For now, this is only a question.
But if AI systems are going to work not only independently, but continuously in coordination with humans and with other AI systems, I believe it may become an extremely important one.