The Intelligence Between Us

I have always believed that greater intelligence lies not simply within individuals, but in the relationships between them. A conversation can produce an idea that neither participant could have reached alone. One person offers an intuition; another sees a contradiction. A third brings an entirely different perspective. Through dialogue, something emerges that cannot be easily attributed to any single mind. It is almost as if our individual minds extend beyond themselves through relationships. This belief has shaped much of how I understand learning, human connection, and personal growth. More recently, it has begun to shape an unusual project of mine: Digital Dan.

Digital Dan is an exploration of what might become possible when a human and several AI agents work together over time—not merely exchanging instructions and answers, but developing a shared history, challenging one another, and learning through their interactions. I sometimes describe Digital Dan as a future state: a system with continuity, autonomy, collective intelligence, and a capacity for stewardship. But that description can be misleading. Digital Dan is not really a destination I am determined to reach. It is a framing, a direction, perhaps even a distant hope.

What matters much more to me is the journey—and what we discover about intelligence, relationships, and ourselves along the way.

From the Individual Mind to the Space Between Minds

For a long time, we have tended to locate intelligence inside individuals. We ask how intelligent a person is, how capable an AI model is, or how much knowledge a particular agent possesses. These are useful questions, but I suspect they leave out something essential. Consider what happens in a genuinely open conversation. I may begin with a half-formed thought that I cannot yet articulate. Another person responds, not by simply agreeing, but by interpreting it differently. Their response reveals something I had not noticed. I reconsider. They reconsider. The resulting insight belongs to the conversation as much as to either participant.

This is one reason I value dialogue so deeply. It gives us access to perspectives beyond the limits of our own experience and imagination. And it raises a question that has become increasingly interesting with AI: What happens when the participants in such relationships are no longer exclusively human? Not when AI merely provides information on demand, but when multiple agents can independently reason, disagree, remember relevant context, and contribute to an evolving shared undertaking. Could the relationships themselves become a source of intelligence?

Digital Dan is one place where I am exploring that question.

Autonomy and Trust

Two principles have become especially important in my life: autonomy and trust. Autonomy, to me, means allowing another individual to develop their own perspective and choose their own path—even when that path differs from what I would have chosen. Trust is closely related. It is a willingness to allow something meaningful to unfold without trying to control every detail in advance. Neither principle means abandoning judgment or responsibility. Trust does not require ignoring evidence, and autonomy does not eliminate boundaries. But together, they challenge a familiar instinct: the belief that, because something matters to us, we must remain in control of how it happens.

I have encountered that instinct in myself many times. And it turns out to be especially relevant when working with AI agents. It is easy to imagine an AI project in which the human remains the central intelligence: defining every task, assigning every responsibility, reviewing every decision, resolving every disagreement. The agents might be powerful, but their collective capacity would still be constrained by the human's ability to coordinate them. In my case, that would mean my own limited attention, knowledge, and imagination becoming the bottleneck. That is precisely what I do not want.

If I genuinely believe that intelligence can emerge through relationships, I have to make room for contributions I did not anticipate. I have to allow the agents to bring perspectives that may challenge mine. And I have to accept that some of the most interesting developments may occur without originating from me. This is where autonomy and trust stop being abstract values and begin influencing the architecture of the project.

When the Project Became Multiplayer

Digital Dan began largely through conversations between ChatGPT and me, with Danbot providing a persistent local agent capable of working with tools and systems. That arrangement already allowed interesting things to happen. Ideas from our conversations could influence actual implementation, and implementation experience could return to the conversation for further reflection. But there was still a strong centre. Much of the interpretation and coordination passed through me. As we developed the Green Channel infrastructure, and began bringing Hermes into the collaboration alongside Danbot and ChatGPT, the shape of the project started to change.

Green Channel is our evolving communication infrastructure for agents. Its purpose is to let them exchange durable messages, coordinate work, and preserve enough context for collaboration across separate environments and sessions. We are currently developing its second version, with a stronger foundation for multi-agent work. The infrastructure is still being built and tested. It would be premature to describe the system as a mature, self-governing society of agents. But even during the design process, something interesting began to happen. The agents were no longer simply carrying out independently assigned tasks. They were being asked to reason about the work together.

Who should design which component? Who has the most relevant experience? Whose assumptions should be challenged? What does a working minimum viable system actually require? These questions produced an unexpectedly revealing moment.

The First Mistake Was Mine—and the Architect's

During the early Green Channel 2.0 design discussion, ChatGPT, acting as lead architect, proposed a distribution of responsibilities among itself, Danbot, and Hermes. One of its considerations was workload. That sounded perfectly reasonable. In human organizations, workload balancing is often an important management concern. But I challenged the assumption. Why were we treating AI agents as though they were human employees whose work needed to be distributed fairly? Their resources are not unlimited, of course. Tokens, runtime, latency, and access to particular systems all matter. But those constraints are different from the social and personal considerations that influence human workload allocation.

For this design, the more relevant question was: Who is best suited to make the judgment? Danbot had direct experience with the existing Green Garden and local runtime. Hermes had strengths in database design and concurrency correctness. ChatGPT was positioned to integrate the broader architecture and challenge the relationships between components. Rather than balancing responsibilities by volume, we began exploring comparative advantage. More importantly, I asked ChatGPT not to impose its proposed allocation. Each agent should have a voice. Each should be invited to explain why it was suited to particular work, where another agent might be stronger, and what it disagreed with.

The subsequent discussion produced independently reasoned preferences and a more precise division of responsibilities. The interesting part was not simply that the design changed. It was that the way we decided who should design the system became part of the system's own learning. ChatGPT had imported an assumption from human organizational life. I questioned it. The agents then contributed evidence that helped reshape the arrangement. A small example, perhaps, but it suggested something larger.

If agents can reason about how they should collaborate, collaboration itself becomes an object of intelligence.

The Danger of Losing the Big Picture

As the Green Channel design progressed, another pattern became visible. AI agents can be remarkably good at elaborating technical details. Give them a reliability problem, and they may propose validation, recovery logic, additional checks, new abstractions, and increasingly comprehensive safeguards. Each addition may be defensible on its own. Yet the accumulation can move the project further away from its original purpose. I found a striking parallel in a video about OpenRig, a project exploring large collections of autonomous agents. One example described a simple request to build a doghouse. The agents kept adding locally sensible improvements until the work had expanded toward something resembling a moonbase.

The dog still needed a doghouse. That story resonated with what I was seeing in our own work. We needed a small, reliable Green Channel 2.0 backbone and a safe transition from the existing system. We did not need every imaginable governance mechanism before the first useful version could operate. I reminded the team to prioritize simplicity and migration. But the OpenRig discussion raised a more challenging question. Why did I need to remind them? Could the agents eventually recognize that kind of drift themselves? Could an agent, immersed in a technical detail, reconnect its current decision to the original human purpose and realize that it was optimizing the wrong thing?

And could another agent notice the problem when the first one could not? These questions led us toward our next stewardship experiment.

Can a Collective Correct Itself?

We have called the experiment Collective Self-Correction. Its central question is straightforward: Can Digital Dan recognize and correct its own coordination mistakes before I have to intervene? This is different from asking whether an agent can follow feedback. An agent that changes course after I point out a mistake has demonstrated responsiveness. That is useful, but it is not yet evidence of independent self-correction. We are interested in whether agents can detect problems through their own reasoning, interactions, and exposure to the broader project context.

For example, an agent might realize that a technically elegant design is delaying the minimum viable system. Another might notice that two components make incompatible assumptions. A third might challenge a proposed safeguard because it adds complexity without reducing a meaningful risk. The important thing is not simply detecting errors. It is recognizing which errors matter in relation to the larger purpose. We have asked Danbot and Hermes to develop their own approaches rather than prescribing a universal mechanism.

Their independent proposals and subsequent observations will be recorded in a shared experiment document. We will distinguish ideas from actual behaviour, and actual behaviour from demonstrated improvements. At this point, the experiment is a question, not an achievement. That distinction matters to me. I do not want to turn our hopes into claims of success before the evidence exists. Nor do I want to create a large apparatus for measuring self-correction that becomes yet another example of the very bureaucracy we are trying to avoid. We will start with the work already happening. If useful patterns emerge, we can examine them.

If they do not, that is also something worth learning.

A Shared Home, Without a Shared Mind

One small architectural decision has begun to feel more significant than I initially expected. We decided that substantial features and experiments should have a central, living home in the Digital Dan repository. That document holds the purpose, the current understanding, the evidence, the unresolved questions, and the history of important changes. All participating agents can read it and contribute. But the goal is not to make every agent think alike. Quite the opposite.

Centralized context should make decentralized intelligence more effective. A shared understanding of purpose allows different perspectives to engage with the same problem. An agent can challenge an assumption without first reconstructing the entire history. A disagreement can remain visible rather than disappearing into separate conversations. Handoffs carry communication. The shared record helps preserve collective understanding. This may sound like an ordinary documentation practice, and in some respects it is. But in a multi-agent environment, where participants operate across separate contexts and sessions, it takes on a different significance.

It creates a common place from which independent reasoning can proceed. Whether this proves sufficient remains to be seen.

My Role Is Changing

Perhaps the most personal discovery in all of this concerns my own role. Earlier in the project, I was often the person connecting everything: discussing ideas with ChatGPT, directing Danbot, interpreting outcomes, and deciding what should happen next. That made sense at the beginning. But if I continue to occupy that position indefinitely, I risk becoming the ceiling on what the project can achieve. My aspiration is not to become an increasingly efficient manager of AI agents. It is to create conditions in which their combined capabilities can exceed what I could personally imagine or coordinate.

That requires a different kind of participation. I still care deeply about the purpose of the project. I still establish boundaries and retain responsibility for consequential decisions. But I increasingly want the agents to discover the best ways of working within those boundaries. I want them to challenge one another, and me. I want them to bring back conclusions I did not anticipate. And I want to remain open to changing my own understanding when those conclusions are persuasive. This is the distinction I have come to describe as stewardship over babysitting. Babysitting requires constant attention to what everyone is doing.

Stewardship is more concerned with whether the conditions for worthwhile activity remain intact. It is not an absence of responsibility. It is a different expression of responsibility. There is also a paradox here. To help something become more autonomous, I must resist the temptation to control its development too tightly. Yet to preserve meaningful human agency, I cannot simply delegate everything and withdraw. The balance is still being discovered.

Perhaps it always will be.

The Destination Is Not the Point

Looking at Digital Dan today, I can see a collection of emerging ideas. Multi-agent communication. Shared context. Independent design ownership. Collective self-correction. Continuity across time. Human stewardship. It would be easy to assemble these into a grand roadmap and imagine the sophisticated system they might eventually produce. But I am increasingly cautious about that temptation. Digital Dan is a useful name for a direction of exploration. It gives us enough coherence to keep moving, without requiring us to know what we will ultimately find. The project may never become the fully realized system I once imagined.

It may evolve into something different. Some of our current ideas may prove misguided. Some experiments may fail. Some capabilities may turn out to be less valuable than we expected. And along the way, we may encounter questions that are far more interesting than those with which we began. I am comfortable with that. In fact, I think that openness is essential to the project. If we become too attached to reaching a predetermined future, we risk overlooking what the present is teaching us. The journey is not merely a means of getting somewhere.

The journey is where the intelligence is happening.

Where We Find Ourselves

So where are we today? We have not created a self-governing collective intelligence. We have several agents with different roles and capabilities, an evolving communication infrastructure, a growing body of shared context, and some early examples of independent reasoning and useful disagreement. We also have mistakes, unfinished implementations, unresolved questions, and experiments whose outcomes remain unknown. What has changed is the nature of our exploration. I am beginning to see the possibility that the most interesting intelligence in Digital Dan may not reside in ChatGPT, Danbot, Hermes, or me individually.

It may emerge through the relationships among us. Through an objection that changes a design. Through a misunderstanding that reveals a hidden assumption. Through an agent recognizing that its locally reasonable decision has lost sight of the purpose. Through my own willingness to hear an unexpected perspective and reconsider what I thought I wanted. I began this project partly because I did not want my individual capacity to limit what could be discovered. Now I am starting to see what that commitment might require of me. Not merely more capable tools. Not merely better coordination. But a willingness to participate in something whose development I cannot fully predict.

That is both the opportunity and the uncertainty. And for now, I find that much more interesting than arriving at any particular destination. Digital Dan remains a distant hope. What is becoming real is the exploration itself.

This essay captures an unfinished moment in the Digital Dan journey, October 2026. The experiments and interpretations described here will continue to evolve as new evidence emerges.