The first generation of DAOs became very good at asking one question: Who gets to decide?
A remarkable amount of infrastructure emerged around that problem. Token voting, delegation, proposal systems, multisigs, treasury controls, councils, quorum requirements, executable governance.
Different projects made different choices, but much of the design space revolved around distributing authority over decisions and shared resources.
It made it possible to coordinate capital and decision rights across people who did not necessarily know each other, share a jurisdiction, or depend on a single organization to enforce the outcome. But after watching DAOs operate for several years, I increasingly wonder whether we concentrated on the wrong layer of the institution; because deciding is only one moment in a much longer institutional process.
A group makes a decision. That decision creates commitments. People or roles act on those commitments. Circumstances change. Some work happens, some does not. Claims are made about what was accomplished. Evidence appears. Responsibilities move. Contributors leave. New people arrive. Money is allocated. An exception quietly becomes normal practice.
Months later, the institution may still have the proposal and the vote. What it often no longer has is a reliable representation of what all of that now means.
The problem may be institutional drift
I have been thinking about this as institutional drift: the gradual divergence between what an institution says, what it does, who is actually authorized to act, what work sustains it, and how its rules evolve in practice.
Consider a simple agreement made during a recurring team meeting. Someone accepts responsibility for a result, circumstances later make the original commitment unrealistic, the team informally accepts a different approach, and eventually everyone behaves according to the new arrangement. Where does the institution now live?
The meeting transcript contains the original conversation. A task manager may contain the outdated commitment. A chat contains the renegotiation. A role document describes who theoretically had authority. The eventual outcome may appear somewhere else entirely.
A human who participated in the whole process may be able to reconstruct it. A newcomer probably cannot.
The institution still exists, but its state has become distributed across artifacts and memories rather than represented anywhere as a coherent whole.
Formal authority and actual authority diverge. Policies survive after the conditions that justified them have disappeared. Important contributors become institutionally invisible. Treasury decisions lose their connection to the assumptions under which funds were allocated.
Eventually, organizations spend increasing amounts of energy reconstructing themselves. They repeat conversations because nobody remembers why a decision was made. They discover commitments only after they have become problems. They debate whether somebody had authority to act after the action already occurred.
What looks like a governance problem can sometimes be a continuity problem.
An institution is more than its decisions
I believe the next important step is not finding a better universal voting mechanism, but learning how to preserve continuity across the rest of the institutional lifecycle.
A rough version of that lifecycle might look like this:
Shared Intent → Coordination → Contribution → Verification → Recognition → Learning → Adaptation
Coordination turns an intention into agreements, roles, decisions, and commitments. Contribution is what actually happens afterward: work, knowledge, relationships, resources, infrastructure, maintenance, care, and the creation of capabilities the institution previously did not have.
Verification asks what we can reasonably say happened. Recognition asks what consequences, economic or otherwise, should follow from that contribution and performance.
Learning asks what the difference between intention and reality tells us. Adaptation changes the institution accordingly.
This could mean modifying an agreement, redefining a role, shifting authority, changing a verification process, reallocating resources, or simply acknowledging that the original assumption was wrong.
The interesting part is that none of these stages can be understood completely in isolation.
A contribution only makes sense in relation to some purpose or commitment. A verification only makes sense in relation to a claim. A payment may only make sense in relation to a verified contribution and an accepted distribution rule.
What connects these stages across time is institutional memory.
That is preserving the relationships that allow the institution to reconstruct why its current state exists.
I think of institutional state as a living graph of institutional relationships. A conversation produces an agreement. That agreement creates a commitment. The commitment is accepted through a role. The role derives its authority from an existing agreement or policy. A contribution fulfills part of that commitment. A claim is made that the work is complete. Evidence supports the claim. A verification evaluates it. A challenge disputes the interpretation. A later decision modifies the original agreement.
None of those objects is particularly revolutionary by itself. The interesting property is that their relationships survive. A future participant could therefore reconstruct not only what the current rule is, but something closer to:
Who acted, through which institutional capacity, under what authority, based on which evidence, and what changed afterward?
This turns institutional history into something closer to an inspectable state. And te implications become more significant when authority, contribution, evidence, and change inhabit the same relational substrate.
Take authority. Most software permissions answer a technical question: can this account perform this action?
Institutional authority asks a different question: why is this actor legitimately allowed to perform this action in this context?
A person may control a multisig key without possessing legitimate authority to exercise it under every circumstance. An AI agent may technically be capable of sending a transaction without having been delegated the institutional authority to decide that the transaction should happen.
A relational institutional state could preserve the provenance of that authority.
Not simply:
Alex approved this.
But:
The Treasury Role approved this; Alex occupied that role at the time; the role received this authority from a particular policy; that policy was established through a previous legitimate process.
The relationship is recoverable when it matters.
Authority becomes something that can be inspected rather than merely assumed.
Institutional truth
Organizations constantly make assertions that quietly become facts.
“Milestone completed.” “Community supported the proposal.” “The contractor delivered.” “The role has authority.”
But an assertion, evidence supporting an assertion, and an institutional decision to accept that assertion are different things. A relational model can preserve those distinctions.
A claim may be supported by evidence. A reviewer may verify it. Another participant may challenge the verification. The institution may ultimately treat the claim as accepted, contested, provisional, or insufficiently evidenced.
That matters increasingly in a world where AI agents can generate institutional records at almost zero marginal cost.
Something entering the institutional graph should not automatically make it true.
This also means that shared institutional state does not have to imply one authoritative account of reality. A healthy state may preserve uncertainty and contain disagreements that have not yet been resolved.
The goal is to make consequential ambiguity visible before somebody builds another decision on top of it.
Contribution becomes easier
There is another consequence I find particularly relevant to DAOs.
Governance has spent enormous attention on who owns and who votes, but comparatively less on the institutional history of who actually made the organization capable of doing what it does.
Contribution is usually scattered across payroll records, task systems, grant reports, forum posts, informal reputation, GitHub histories, and people’s recollections.
A shared institutional graph could make contribution another type of relationship.
A role performed work. That work contributed to a result. Evidence exists. The contribution strengthened a capability. Later, the community may decide whether and how that should matter for recognition or distribution.
This makes it possible to ask interesting questions.
Which capabilities actually sustain this organization? Which depend entirely on one person? Which forms of invisible work repeatedly disappear from the official record? Who created value long before the institution had enough liquidity to compensate for it?
Those are governance questions too.
Organizations have spent decades trying to structure roles, responsibilities, processes, decisions, records, and knowledge.
The common failure mode is that the structure itself becomes work. Someone has to maintain it, and eventually the formalization burden exceeds the value of the additional legibility.
This is where AI may alter the tradeoff.
What if people could obtain the benefits of structured institutional state without having to perform most of the formalization themselves?
Imagine a normal meeting. Someone simply says:
“We agreed that the communications role will publish the report next Friday.”
An agent proposes that an agreement exists, identifies the related commitment, associates it with the appropriate role, and connects it to the relevant context.
A human reviews the proposed state.
Later someone says:
“Friday is no longer realistic. We can do Monday because the final data arrived late.”
The system proposes a change rather than silently overwriting the original commitment. The participants still speak normally. The ontology remains largely invisible.
That, to me, is the interesting technological hypothesis. Not that AI should govern an organization. Not that an LLM should decide what is true.
But that AI might dramatically reduce the interface cost of institutional formalization.
An AI might be able to infer that a commitment has changed. That does not mean it should be allowed to change institutional state without confirmation. It may be able to determine that submitted evidence appears sufficient. That does not mean it necessarily has authority to approve payment.
I find it useful to keep five things separate:
Observation → Interpretation → Recommendation → Authority → Execution
Different actors can legitimately inhabit different points in that chain.
An AI agent could observe and structure information. A human or role could authorize a consequential change. A smart contract could execute the resulting settlement. An agent might have delegated authority to perform a narrow operational transition autonomously.
The architecture needs to preserve enough state to know who is allowed to do what, and why.
This is still only a hypothesis
I would start by asking:
Can an AI-assisted workflow turn ordinary coordination conversations into a small, human-validated institutional graph that is accurate and useful enough to improve continuity, without requiring participants to understand the underlying ontology?
That is the experiment I would like to contribute to the v2 DAO conversation.
And it can fail.
The first thing to measure would be fidelity. Does the system correctly identify agreements, commitments, responsible roles, changes, and relevant evidence? Or does it hallucinate institutional structure that participants never intended?
The second is correction burden. How much human work is needed before the proposed state becomes trustworthy enough to use? If the agent saves ten minutes of extraction and creates twenty minutes of correction work, the experiment is failing on its own terms.
A third measure is actual downstream utility. Does the graph help recover a forgotten commitment? Surface a blocker earlier? Clarify why someone had authority? Reduce the time required to reconstruct a previous decision? Prevent an outdated agreement from quietly surviving into another cycle?
And the fourth might be the most important: Can participants receive those benefits without learning the ontology?
If people need to understand nodes, edges, schemas, provenance chains, and state transitions before the system becomes useful, we have probably moved administrative work rather than eliminated it.
The relevant baseline is: How does this group coordinate today?
Take an organization with an existing recurring meeting. Observe what happens with ordinary notes, task managers, chats, and human memory.
How many meaningful commitments survive from one meeting to the next? How many are silently abandoned? How much time is spent reconstructing what had previously been agreed? How often do participants disagree about the state of a commitment?
Then introduce the smallest possible institutional-state workflow and compare. That experiment could tell us whether the basic design tradeoff is feasible.
Why I think this is worth testing now
The v2 DAO conversation is currently asking important questions about legitimacy, organizational form, modular governance, and what should replace some of the assumptions inherited from the previous cycle.
I would like to add one hypothesis to that conversation: Perhaps distributed institutions need a continuity layer.
A layer capable of remembering how intention became agreement, how agreement became action, how action produced claims and evidence, how value was recognized, and how the institution changed afterward.
If that layer can be maintained cheaply enough, many other institutional capabilities become easier to compose. Different decision mechanisms can coexist because their outputs enter the same institutional state.
Humans and agents can inhabit roles without authority becoming synonymous with identity. Contribution can become visible without automatically becoming compensation. Smart contracts can execute sufficiently specified decisions without pretending to resolve ambiguous social questions. And organizations can experiment while retaining the evidence necessary to learn from what they tried.
The target should be minimum sufficient institutional legibility.
Enough structure to preserve consequential relationships. Enough provenance to understand why the current state exists. Enough contestability to correct it. Enough history to learn. And little enough administrative burden that people can continue doing the work the institution exists to support.
Toward v2 DAOs
The first generation of DAOs asked how decision-making power could be distributed without relying entirely on traditional organizational structures.
That experiment opened a real design space. Perhaps the next question sits one layer deeper.
How can institutional capabilities be distributed without allowing the institution itself to fragment?
I believe a living graph of institutional relationships is one possible answer.
A graph as a continuity layer: a way for an institution to retain an inspectable relationship between who acted, through which institutional capacity, under what authority, based on what evidence, and what changed afterward.
If AI can maintain that state while humans continue interacting in ordinary language, we may have a feasible new primitive for distributed organization. If it cannot, we should learn that before building an elaborate architecture on top of it.
Lets learn how to build institutions that can remember what they tried, produce evidence about what happened, and change when the evidence says they were wrong.
Can you explain the need for blockchain in the core idea? The essay makes a strong case for a provenance-aware knowledge graph plus AI assistance, but it's not clear why that graph has to be on-chain. Smart contracts make sense for narrow execution once a decision is authoritative, but the difficult parts like interpretation, legitimacy, and disputed evidence remain social.