The right to
computational assembly

The capacity the public has been missing

A household can have excellent information and still negotiate badly. Knowing that a contract is expensive does not create an alternative supplier. Understanding an unfair term does not provide the time, legal support, or bargaining position needed to change it.

Organizations overcome these problems by maintaining capacity between decisions. Their records persist. Staff specialize. Budgets pay for investigation and follow-through. An individual usually has to reconstruct the relevant knowledge whenever a problem returns.

Personal agents could reduce some of that burden. They could remember contractual conditions, monitor deadlines, compare terms, and prepare actions within a person’s instructions. Cooperation between agents could extend that capacity to groups whose members share a particular interest but cannot sustain a permanent organization around it.

The electricity coalition would need more than a clever model. It would need permission to discover other buyers, a way to verify mandates, rules for handling disagreement, and someone answerable when a transaction fails. It would also need to survive the departure of its software supplier.

These are institutional requirements. Improvements in model intelligence alone will not meet them.

That distinction matters for policy. A society can make advanced models widely available while leaving the practical means of collective action under a few providers’ control. People might be free to consult an assistant, yet unable to use it to negotiate with the institutions setting their terms.

The public interest therefore includes the conditions under which personal agents can work together. Access to an individual service is only part of the question.

How safety requirements can favor concentration

Artificial intelligence systems capable of taking consequential actions require oversight. A regulator needs to know who operates a service, what authority it has, and how intervention would work. An insurer needs enough evidence to assess exposure. A counterparty needs confidence that an agent can make the commitment it offers.

Concentration can emerge from the way those requirements fit together.

Suppose an agent operator must use an accredited identity service. The identity service accepts only a small set of certified execution environments. Certification requires an extensive evaluation whose cost is largely fixed. Insurers then price coverage around those environments, and major transaction networks accept only insured operators.

No single requirement expressly prohibits a cooperative. Taken together, they may leave it with one affordable option: operate inside an incumbent’s service and accept that service’s terms.

This is a hypothetical mechanism, not a description of a universal licensing regime already in force. It shows why each rule needs to be evaluated within the system it helps create. A requirement can be reasonable in isolation and exclusionary in combination.

There is an existing infrastructure problem to consider. The Federal Trade Commission’s January 2025 staff report on major cloud and artificial intelligence partnerships identified potential competition concerns involving access to computing resources, contractual and technical switching costs, and the sharing of sensitive information. These were findings about relationships and their possible effects, not a determination that the partnerships were unlawful. They establish a concrete basis for examining dependency before adding further requirements to the same infrastructure. [1]

The European Union’s Artificial Intelligence Act illustrates a different distinction. Article 51 provides criteria for classifying general-purpose models as presenting systemic risk, including a presumption associated with training computation above 10²⁵ floating-point operations. That classification concerns models. It does not itself establish a general prohibition on citizens coordinating through agents. Treating it as evidence of such a prohibition would misstate the law. [2]

The policy concern is prospective: obligations appropriate to developing a powerful model should not automatically become obligations imposed on every group using a bounded application of one. Developers, service operators, and participating households control different things. Their responsibilities should reflect those differences.

This argument does not depend on claims that executives manufactured public concern or privately agreed on a common outcome. Such claims require their own evidence. The relevant institutional test is whether the resulting rules leave independent groups a workable route to compliance.

What computational assembly would protect

Computational assembly is the voluntary cooperation of software agents acting under mandates from people or legitimately constituted organizations. In the Commonsent Dictionary, an agent acting for one person under such a mandate is a personal advocate agent.

The protected interest belongs to the people represented. Agents acquire no independent political standing from participating. Running a thousand agents must not give someone a thousand votes, manufacture a thousand citizens, or create authority that no person granted.

For a participant, the proposed right would have several practical consequences. They could choose a representative, define its authority, and permit it to cooperate with representatives supplied by other providers. They could inspect consequential commitments and withdraw permission for future action. They could also change providers without automatically losing membership in the association.

A workable formulation is:

People should presumptively be free to delegate lawful activity to computational agents and authorize those agents to cooperate. Restrictions should address identifiable risks, protect affected parties, and preserve a practical route to compliant participation.

The qualification matters. A purchasing mandate does not authorize intrusion into a supplier’s systems. Membership in a civic group does not authorize automated harassment. A coalition cannot consent on behalf of people outside it merely because its participants agree.

Nor does the right promise escape from commitments already made. Revoking an agent’s authority should stop new commitments; it does not necessarily cancel a valid contract concluded before revocation. A system must make that boundary understandable before asking anyone to delegate.

The principle would also be incomplete without protection against compulsory participation. A person must be able to decline an association, reject an expanded mandate, and contest an agent’s interpretation of their instructions. Otherwise, collective agency could become another source of pressure on the individual.

The difficult work lies in making these conditions enforceable when participation is large, software is imperfect, and people are busy.

Safety has to account for the group

The original attraction of a distributed network is also a source of risk: many modest actions can combine into a large effect.

Thousands of authorized agents might overload a public service, react simultaneously to a false signal, or concentrate purchases in ways that damage a market. A participant’s individual spending limit says little about the aggregate exposure of a coalition. Shared software can reproduce the same error across otherwise independent members.

For that reason, regulating harmful outcomes alone would be inadequate. Some harms cannot be repaired after the event. Preventive requirements may be justified by capability, deployment conditions, or credible evidence of systemic risk.

The regulatory task is to identify the relevant mechanism. The word swarm provides little help. It can describe coordinated malware, a collection of customer-service tools, or households seeking a joint offer. Those systems differ in their permissions, targets, external effects, and means of intervention.

For the electricity example, safeguards should follow the authority actually being exercised:

These controls address different failures. Combining them into a single “trusted agent” badge would obscure the distinctions.

Technical enforcement also has limits. A signed instruction can establish that a key authorized something; it cannot establish that the person understood the consequences or that the action was fair to others. A transaction log can help reconstruct a failure; it cannot recover money that has already become irretrievable.

Independent review, accessible complaints, and legal responsibility remain necessary. Some activities may require licensed operators or stronger supervision. Where those requirements are justified, policy should examine whether shared services, public evaluation facilities, or other arrangements can make compliance accessible without lowering the standard.

The Artificial Intelligence Risk Management Framework from the National Institute of Standards and Technology supplies a useful discipline here: risks are governed, mapped, measured, and managed throughout a system’s life. It does not establish that federation is safe or endorse this proposed right. It supports assessing the system in context rather than assuming that an organizational label resolves the risk. [3]

A federated system should have to show that its controls work. Claims of decentralization are insufficient evidence.

Collective power remains accountable

A right to computational assembly would protect the ability to form an association. It would not settle the legality of everything the association does.

Purchasing groups can improve terms for participants. They can also acquire enough power to exclude suppliers, pressure dependent businesses, or restrict competition. An agent network representing buyers deserves no automatic exemption because its members describe themselves as a community.

Existing competition analysis offers distinctions worth preserving. The European Commission’s horizontal cooperation guidelines distinguish genuine joint purchasing from buyer cartels, while recognizing that joint purchasing can itself create competition concerns. This is a specific legal framework for cooperation between undertakings, not a blanket authorization for every household coalition. It nevertheless demonstrates why “buyers acting together” is too broad a category to determine legality. [4]

For an agent-mediated arrangement, scrutiny should reach the people and organizations that actually control it. Ten thousand nominally personal agents could all be subject to the same provider’s commercial instructions. A federation with many operators could still depend on one ranking service that decides which suppliers receive demand.

The inquiry should therefore examine mandates, shared control, information exchange, market effects, and the ability to refuse collective decisions. Counting agents or servers will not answer those questions.

Responsibility must also be allocated before deployment. Participants should be answerable for conduct they knowingly authorize. Developers and operators should remain answerable for failures within their control. A household should not become the default bearer of every downstream loss simply because its name appears on a mandate.

Distributed operation makes these arrangements harder to specify. That is work the design must undertake, including a credible route for injured outsiders to obtain a remedy. An accountability system that only its members can use is incomplete.

Where control accumulates

The most restrictive system need not announce a ban on independent agents. Dependence can develop through ordinary product decisions.

A provider might hold the only usable copy of a person’s history. Its identity service might be the only one accepted by counterparties. Its reputation score might determine access to offers, while the evidence behind the score remains inaccessible. A coalition might be exportable in principle but unable to move its active mandates or contractual records.

Each dependency makes leaving harder. Together, they can give a supplier effective control over an association it does not formally govern.

Interoperability must therefore cover the relationships that matter. Sending messages between two agents is a beginning. Participants also need a way to preserve valid authority, retrieve records, settle outstanding obligations, and continue working with one another after a provider changes.

There are legitimate limits. Personal data concerning other members cannot simply be copied wherever one participant chooses. Fraud controls cannot be discarded during migration. An agent’s performance may change when its underlying model changes. Portability requires rules for these cases and tests that expose where continuity fails.

The same scrutiny belongs inside Commonsent. A participant-owned organization could still concentrate power in a technical team, an identity registry, or the people controlling a shared treasury. Formal ownership would offer little protection if members could not replace an operator or challenge a decision.

This is where Authority Before Infrastructure provides a useful architectural foundation. It separates participant authority from procured computational capacity. Applied here, that means a service provider may perform work for an association without acquiring the right to redefine its membership, expand its mandates, or prevent its departure. [5]

That separation must survive an actual provider failure. Otherwise it remains a contractual aspiration.

What Commonsent would have to demonstrate

The proposed right gives Commonsent a demanding design obligation. A network advocating computational assembly should be able to demonstrate meaningful participant control in a limited deployment before claiming it at societal scale.

A first pilot could organize a bounded purchasing exercise. It would not need unrestricted autonomy or thousands of participants. It would need enough variation to test disagreement, withdrawal, provider substitution, and failure.

Participants would begin by authorizing specific activity: seek offers for a defined service, use specified information, observe stated constraints, and return for approval at an agreed point. The system would record those instructions in a form that both people and software could inspect.

Coalition formation would then require more than finding similar preferences. Participants would need to know how offers would be evaluated, who could negotiate, how operating costs would be paid, and what would happen if only some members accepted the result. Differences would remain visible. A preference held by most participants would not silently become permission granted by everyone.

Before commitments were made, the system would have to check that the relevant mandates were still valid. If an agent exceeded its authority, there would need to be an identified response process and a responsible party. A dispute could not be resolved solely by asking the same software to explain why it thought the action was appropriate.

The pilot should deliberately interrupt its own infrastructure. Replace the model provider. Disable a coordinator. Attempt a migration while offers are pending. Test whether revoked permissions remain revoked across the network. These exercises would reveal dependencies that a successful demonstration under normal conditions might conceal.

The resulting evaluation should report more than savings. It should measure the effort required to participate, the cost of compliance, unauthorized commitments, recovery times, and the time needed to move to another operator. It should examine whether people with limited technical knowledge could understand and change their mandates.

Negative findings would be consequential. If compliance consumed all the economic benefit, the proposed service would not yet work. If participants could exit only by abandoning their history or accepting substantial losses, its claim to independence would be weak. If distributed controls could not contain a demonstrated risk, that deployment would need stronger restrictions or a different design.

These are conditions under which the proposal can be tested and revised. They prevent a desirable political objective from standing in for operational evidence.

A test for rules governing agents

Policy affecting collective agents should publish an assessment of its effects on lawful association alongside its safety analysis. That assessment should answer five questions.

What failure is the rule intended to prevent?

Identify the capability, deployment condition, or observed behavior creating the danger. Explain how the requirement would reduce it.

Who can comply in practice?

Estimate the cost and burden for a small operator, cooperative, municipality, and large company. Include the combined effect of related requirements.

Where does the rule place control?

Identify any provider, registry, evaluator, or communications service that becomes indispensable. State how its decisions can be challenged.

Could another arrangement provide adequate protection?

Examine bounded permissions, transaction controls, independent audits, shared compliance services, and restricted deployment. Where those measures are insufficient, explain why.

How will the rule be reviewed?

Specify the evidence that would justify tightening, revising, or removing it. Monitor whether compliant independent operators can actually enter and remain in the market.

An unequal effect would require investigation rather than automatically invalidate a rule. Some activities are costly to supervise because the potential harm is severe. The relevant question is whether that burden follows from the risk or from avoidable dependence on a favored institutional form.

The assessment should include people who would have to operate under the rule. Frontier laboratories have essential expertise, but they cannot supply every relevant perspective. Independent developers, consumer groups, security researchers, cooperatives, and affected communities need resources to test proposals and submit evidence. An invitation to comment does little for an organization unable to afford the analysis.

This would make competition and civil agency part of regulatory design from the beginning, when requirements can still be changed.

The terms on which people can act together

The electricity coalition is a modest example. Its members are trying to obtain better terms for a recurring household expense. They need competent representation, protection against mistakes, and a way to work together without surrendering more authority than the task requires.

A workable regime would allow them to do that. It would require the coalition to demonstrate valid mandates, handle data responsibly, honor commitments, and answer for harm. It would also leave them a practical choice of operators and a way to continue if one withdrew.

A restrictive regime could offer the same households capable personal assistants while reserving collective negotiation for institutions able to afford a permanent compliance apparatus. The public would receive help making decisions within existing terms, while the ability to bargain over those terms remained elsewhere.

That is the possibility this proposal addresses.

Computational assembly deserves protection because people will increasingly exercise ordinary agency through software. Their ability to cooperate should survive that change of medium. Establishing the principle now would give lawmakers, standards bodies, and system designers a concrete obligation: preserve a workable route for people to authorize joint action, govern it, and leave it.

The test will be whether the households can complete the negotiation on terms they understand, with responsibilities they can meet, and still control their association when the contract is signed.

Evidence and related work

1. Federal Trade Commission: artificial intelligence partnerships and investments studyJanuary 2025. Evidence on potential competition effects of cloud and artificial intelligence partnerships.
2. European Union: Artificial Intelligence Act, Article 51Regulation (European Union) 2024/1689. Classification of general-purpose artificial intelligence models with systemic risk.
3. National Institute of Standards and Technology: Artificial Intelligence Risk Management FrameworkVersion 1.0. Contextual risk management through Govern, Map, Measure and Manage.
4. European Commission: horizontal cooperation guidelines2023, Chapter 4. Joint purchasing arrangements and buyer cartels.
5. Commonsent Lab: Authority Before InfrastructureAugust 2026. Participant authority, provider substitution and progressive infrastructure independence.