Institutions of intelligence
Most AI systems still preserve the conceptual form of a single interlocutor: one model, one context, one answer, one apparent point of view. Even systems that invoke several agents often return a larger monologue or a temporary theater of personalities.
I am interested in another possibility: AI not merely as an assistant, a tool, or a crowd of simulated voices, but as a participant in durable institutions of inquiry and action built around human purposes.
Human beings routinely exceed the limits of individual cognition through institutions. A court, university, laboratory, editorial process, company, archive, or intellectual tradition can preserve memory, distribute attention, sustain disagreement, and support forms of judgment that no participant could reliably reproduce alone. Institutions are not merely containers for intelligent people. They are themselves technologies for cognition and agency.
They are also fallible technologies. Bureaucracy can replace purpose. Consensus can erase truth. Procedure can become theater. Authority can become detached from responsibility. Institutional form does not make intelligence virtuous. It makes consequential relationships among knowledge, judgment, memory, and action explicit enough to examine.
Language models make cognitive labor available with unusual abundance and flexibility. But abundance alone can produce noise, imitation, premature consensus, and hidden transfers of authority. The interesting question is not how to make more agents speak. It is whether their labor can acquire a form that is rigorous, cumulative, inspectable, and answerable to a person.
I call this a constitutional knowledge process. Its sources of authority, forms of contribution, treatment of evidence and disagreement, relationship to memory, and power to authorize change are not left implicit in a model’s performance. The constitution need not imitate a parliament or a court. It is the governed shape through which a particular kind of thought becomes possible.
The next important unit of AI design may not be the prompt or the agent, but the institution.
Externalizing cognition is not new. People think through writing, mathematical notation, conversation, archives, professions, disciplines, and institutions. Each allows part of cognition to persist outside one person’s working memory and return in a form that can alter later thought.
The new possibility is to externalize much more of the active work around judgment: searching, checking, holding alternatives in view, reconstructing arguments, connecting material across time, and coordinating complex action. The purpose is not to avoid thinking. It is to spend less human attention on the combinatorial burden that so often prevents deeper thought.
That possibility carries a real danger. A system that always supplies the synthesis may cause the capacities through which its user understands, judges, and authors a life to atrophy. A system that recalls every settled and abandoned belief without preserving the difference may become an echo chamber with infinite patience. A system that recommends action may quietly become the real seat of practical judgment even when the human retains a ceremonial final click.
I therefore treat human sovereignty as a design boundary rather than a disclaimer. These systems may enlarge the field of perception, preserve more possibilities, perform more verification, and coordinate more work. The human remains responsible for understanding, judgment, assent, changes in value, and final authorization. The system must preserve that responsibility without protecting the human from surprise.
Externalize cognitive labor without externalizing human sovereignty.
Convoco names the research and engineering program that contains Franklin and Abelard. It asks how AI can participate in durable institutions of inquiry and action without displacing human authorship. Franklin and Abelard approach that problem from different directions.
Franklin approaches the problem through agency. It asks how a human intention can become ambitious, coordinated work without dissolving authorship or hiding responsibility.
The purpose is not merely to automate tasks. It is to let one person remain meaningfully sovereign over work whose scale would otherwise exceed that person’s attention.
Intention becomes action.
Abelard approaches the problem through knowledge and inquiry. It asks how sources, arguments, intellectual traditions, personal writing, disagreement, and historical change can become active in present thought without collapsing into a search index, model persona, or static belief profile.
The purpose is not merely to answer questions. It is to make more of the world—and more of a person’s own intellectual life—available to thought while leaving that person more capable of understanding and judging.
Encounter becomes understanding.
Structured deliberation—giving conflicting perspectives enough independence to confront one another—offers one way into this problem. It may improve a model’s answer, but the debate disappears when the context closes. The larger question is what would allow organized AI cognition to become a durable, evaluable, and governable institution rather than another transient conversation.
Several distinctions keep the project honest.
Convoco is not a theory that additional agents automatically produce better judgment. A chorus of simulated voices can manufacture the appearance of plurality while sharing the same blind spots. Procedure can amplify error as readily as insight.
Convoco is not a generic workflow system. Connecting model calls can coordinate activity, but coordination alone does not establish what counts as evidence, what deserves independence, what may become memory, or who has the authority to make a conclusion operative.
Convoco is not an attempt to build a digital clone of me. A system that predicts what I would say is different from a body of thought in which I have deliberately decided what may speak for me.
Convoco is not an attempt to transfer moral, political, or practical authority to machines. The systems may inform future judgment and action. They do not acquire the right to make those judgments or take those actions on my behalf.
The ideas become clearer in use. These stories describe the transformations I want the systems to make possible.
I have a product idea whose implications exceed what I can hold in working memory. It touches product judgment, architecture, implementation, evaluation, and deployment. Ordinarily I would either narrow the ambition or spend most of my attention maintaining coordination rather than thinking about the product.
With Franklin, I develop the intention at the level where my judgment matters. The surrounding work becomes organized, examined, and carried forward without requiring me to impersonate an entire company in my own head. When a choice changes the purpose, values, or acceptable result, it returns to me. When the work can be judged from what I have already authorized, it proceeds.
I remain the author, but authorship no longer requires personally performing or supervising every cognitive act.
I read an editorial whose argument feels persuasive and incomplete at the same time. It blends reported facts, causal claims, moral language, historical analogy, and political framing. A chat model can summarize it or offer a counterargument, but I would still need to verify the answer and reconstruct how it relates to everything else I know.
With Abelard, the material becomes an inquiry rather than another exchange of confident prose. The relevant facts can be checked. The argument can be reconstructed at its strongest. Alternative explanations and deeper principles can enter. My prior thought can be consulted without being mistaken for external evidence. The resulting account can preserve genuine division or uncertainty when the evidence does.
I receive more than an answer. I return to the subject with sharper perception and a greater ability to discuss it myself.
When does AI-assisted reasoning enlarge agency, and when does it weaken the capacities through which agency is exercised?
What would it mean for an AI system to represent an institution rather than imitate its voice?
How can disagreement remain real without manufacturing opposition for its own sake?
How should a body of knowledge change without rewriting the history of what it previously held?
What should machines remember, and what should require renewed human attention?
How can a system faithfully represent what I believe while continuing to expose me to what I would not have thought to ask?
How can a person remain the author of work produced through cognition they did not personally perform?
These are not questions surrounding the software. They are the work.