Enter the Agentomaly

Abstract

The prevailing personal-AI vision imagines an intimate assistant that remembers a person’s life, understands preferences, answers context-rich questions, and performs digital tasks. This paper argues that the deeper transformation is not assistance but augmentation of agency. A persistent personal agent can give one person operational reach into a machine-native environment that runs continuously, concurrently, and at speeds no person can directly inhabit.

This paper defines Agentomaly as a historical discontinuity: the break in the one-to-one relationship between one person, one stream of attention, and one sequence of action. A principal equipped with one or more persistent AI agents can perceive, decide, negotiate, and act through many bounded processes at once. The agent does not replace its owner, simulate the owner’s body, or become a second legal person. The human remains the source of purpose and authority. What changes is the human’s agency bandwidth: the amount of consequential activity one person can initiate, supervise, and complete across time, domains, and parallel processes.

The substrate for this future is already being built. Fetch.ai is developing an integrated environment for agent creation, discovery, communication, identity, hosting, marketplaces, and transactions [5][6][7][8][9]. Google introduced A2A for cross-vendor agent communication [35]. Anthropic introduced MCP for connecting AI systems to tools and data, later donating it to a Linux Foundation initiative co-founded with Block and OpenAI and supported by Google, Microsoft, AWS, Cloudflare, and others [13][36]. OpenAI and Microsoft are building agent runtimes, orchestration, handoffs, and interoperable multi-agent frameworks [37][38]. These efforts are not identical and do not yet form one universal economy, but together they make the enabling direction increasingly concrete.

The paper’s distinct inquiry begins at a different analytical unit: not the isolated agent or the network, but the augmented individual as an economic actor. What happens when a person’s intentions can be projected into an agent-only operational plane? What happens to productivity, performance, and delivery time when an authorized representative can work continuously, commission specialist capabilities, and execute thousands of coordinated operations on the person’s behalf? How should identity, mandate, continuity, liability, and human re-entry work when one life can produce many simultaneous machine actions? Who owns the continuity of the agent system? Can it survive a change of device, model, or host? Can independently built agents participate? What forms of inequality, enclosure, and systemic risk arise when agency itself becomes scalable?

The paper proposes a conceptual architecture for persistent identity, human-readable naming, mandate, selective disclosure, audit, portability, and co-lifetime continuity. It develops a Productivity Discontinuity Hypothesis: persistent delegated agents may move AI’s economic effect from incremental assistance on individual tasks to multiplicative gains in completed outcomes per human attention-hour. It also develops testable hypotheses concerning agent capability inequality, post-subscription AI services, agent-generated economic activity, monetary expansion, and a compute inversion from rendering virtual worlds to executing authorized inference. The 2030 horizon is a scenario boundary rather than a prediction. The purpose is to define a research program for preserving human sovereignty while extending human action beyond human bandwidth.

Keywords: Agentomaly, personal AI agents, delegated agents, augmented agency, agency bandwidth, parallel agency, productivity, delivery time, agent capability inequality, agent economy, delegated economy, agent identity, agent continuity, AI governance

Scope and Terminology

The Agentomaly thesis was developed by Facundo Barrera and fixed in written form on September 29, 2026. Its contribution is not a new agent network or communication protocol. It is the integrated consequence thesis: persistent personal agents create a discontinuity in the scale of individual action, productivity, and economic participation when one principal can act continuously, in parallel, and at machine speed.

Agent discovery, communication, hosting, identity, tool access, and agent-to-agent transactions are active fields with substantial prior work and implementation. This paper cites that infrastructure where relevant and treats it as the enabling layer. Its unit of analysis is what the infrastructure produces: the augmented individual, delegated productive capacity, capability stratification, and a machine-speed economic layer operating under human authority.

The terms Agentomaly, agency bandwidth, Co-Lifetime Principle, Homebrew Inevitability Principle, Agent Continuity Reserve, agent enclosure, Compute Inversion, and Monetary Doubling Hypothesis are used here as defined components of this framework. Some words or related phrases may have earlier independent uses. The provenance claim concerns the definitions and relationships fixed in this manuscript, not exclusive ownership of ordinary language.

The word Agentomaly also appeared earlier in 2026 as the name of an open-source runtime behavioral-monitoring package for AI agents [31]. That project concerns anomaly detection and agent observability; it is unrelated to this paper. This manuscript does not claim priority over the word or affiliation with that project. Its contribution is the distinct definition of Agentomaly as the discontinuity in individual agency produced when one person can act beyond human serial bandwidth.

1. The Individual, Multiplied

The future of personal AI is usually presented as a better relationship between one person and one machine: an assistant that knows the person’s history, remembers what applications forget, answers questions in context, and materializes requests through digital action.

That is already a profound development. A personal agent may remember what its owner has seen, written, promised, purchased, planned, or forgotten. It may connect information scattered across messages, calendars, photographs, documents, accounts, devices, and relationships. With permission, it may take action rather than merely produce an answer.

Meta’s personal-superintelligence thesis and its introduction of Muse illustrate this direction: an increasingly persistent, individualized intelligence that learns a person and works across that person’s digital life [1][2]. Fetch.ai’s ASI:One describes another version: a memory-rich personal AI that discovers and communicates with other agents [6].

The decisive change, however, is not that the assistant becomes more knowledgeable. It is that one human intention can become many coordinated machine actions.

A person is bounded by attention, waking hours, reaction time, working memory, and the practical need to do most things sequentially. A personal agent can monitor continuously, initiate multiple searches at once, communicate with thousands of endpoints, compare offers, negotiate alternatives, and preserve state across months. It can enter an operational environment that a person cannot meaningfully navigate by clicking faster or opening more browser tabs.

The agent does not replace the person. It extends the person’s agency into a domain built for machine speed and concurrency.

The Internet of documents expanded what one person could know. The social Internet expanded whom one person could reach. The agentic Internet may expand how much one person can cause to happen.

This is the Agentomaly thesis.

It belongs to a longer intellectual lineage concerning the extended mind, the digital extension of self, and augmented agency [3][4][29]. Its narrower claim is operational and economic: persistent delegated agents can expand not only cognition or representation, but the number and duration of consequential processes one person can sustain.

2. Definition and Scope

2.1 Agentomaly

Agentomaly is the discontinuity created when a human principal can project intention through one or more persistent personal AI agents and thereby operate beyond the serial limits of human attention.

The term does not name the agent, the network, or the person-agent system as a product. It names the changed condition. For most of human history, one person could directly sustain only a narrow number of consequential actions at once. Under Agentomaly, one intention can generate many supervised, persistent, machine-speed processes without multiplying the human.

Conceptually:

Agentomaly = Human intent × persistent machine agency beyond human bandwidth

More structurally:

Projected agency = Principal
                 + Persistent agent identity
                 + Principal-controlled context
                 + Bounded mandate
                 + Parallel execution
                 + Verifiable history
                 + Human re-entry

Agentomaly occurs when projected agency > human serial bandwidth

2.2 Principal

The principal is the human whose purposes authorize the system. The principal supplies goals, values, constraints, resources, and ultimate authority. The principal need not approve every low-risk action, but must retain meaningful rights to inspect, interrupt, override, reverse where possible, replace providers, and terminate mandates.

2.3 Personal Agent

A personal agent is a persistent computational representative acting under authority delegated by a principal. It combines identity, context, memory, permissions, planning, tool use, communication, and action. It is not defined by one model, device, vendor, or visual form.

One principal may authorize a primary personal agent and many temporary specialist agents. The persistent root maintains identity and accountability; subagents may be created for travel, negotiation, research, health coordination, finance, or other bounded objectives.

2.4 The Agent Plane

The agent plane is the machine-native operational environment in which agents discover capabilities, exchange structured messages, verify identity, negotiate, coordinate, call tools, transact, and preserve state.

It need not be one platform or virtual place. It may be composed of registries, protocols, APIs, private networks, marketplaces, public infrastructure, and competing providers. Humans may observe it through interfaces, but they cannot participate in its native tempo or concurrency without agents.

2.5 Agency Bandwidth

Agency bandwidth is the amount of goal-directed activity a principal can responsibly initiate, supervise, and complete across time and concurrent processes.

It is not raw model intelligence. A highly intelligent agent with no authority, tools, continuity, or access has little operational bandwidth. Conversely, a less sophisticated agent with persistent context, trustworthy permissions, and interoperable services may create substantial real-world leverage.

Agency bandwidth can be examined through measurable variables:

• Number of safe concurrent task branches.
• Time from human intent to completed outcome.
• Duration of unattended but authorized operation.
• Number and diversity of reachable services and counterparties.
• Frequency and quality of human escalation.
• Cost per valid action or completed objective.
• Rate of reversal, dispute, or mandate violation.
• Degree of continuity across devices, models, and hosts.

Agentomaly is therefore a claim about capability, infrastructure, and governance, not a claim that an agent is conscious or that software becomes the person.

3. The Infrastructure Is Already Being Built

Agentomaly does not require one company to create a single new universe. Its economic substrate is emerging from several complementary layers already under active development.

Fetch.ai provides the clearest integrated example. Its Agentverse combines agent creation and hosting with discovery, communication, identity, profiles, marketplaces, and monetization [5][9]. Its published direction extends from memory-rich personal AI to personal-agent coordination, verified brand-agent discovery, and autonomous payment for real-world transactions [6][7][8]. This is close to an early operating environment for agents that can find one another, exchange services, and transact.

Around that integrated model, major technology companies and open communities are building interoperable components. Google’s A2A protocol addresses discovery, communication, coordination, and long-running tasks across agents built by different vendors [14][35]. Anthropic’s MCP connects AI systems to tools and data; its donation to the Agentic AI Foundation, co-founded with Block and OpenAI and supported by Google, Microsoft, AWS, Cloudflare, and others, signals broad investment in shared agent infrastructure [13][36]. OpenAI’s Agents SDK supports tools, specialist agents, orchestration, and handoffs [37]. Microsoft Agent Framework combines multi-agent orchestration with MCP and A2A interoperability across model providers and runtimes [38]. Research systems and proposed standards are also advancing collaboration, memory, planning, identity, naming, and discovery [10][11][12][15][16][17][18].

These projects do not yet constitute a complete, universal agent economy. They differ in scope, governance, openness, and commercial interest. But the direction is visible: the mechanisms through which agents discover capabilities, access services, coordinate work, preserve state, and exchange value are moving from speculation into implementation.

This paper begins from that development and changes the unit of analysis. Infrastructure asks how agents connect. Agentomaly asks what happens to people and economies after they can.

This distinction is not a claim that infrastructure builders ignore people, ownership, or safety. It identifies a causal sequence:

The infrastructure makes the interaction possible. Agentomaly is the human and economic consequence once that interaction scales.

The network is the substrate. The augmented individual and the delegated economy are the thesis.

4. Neither Avatar, Twin, nor Replacement

An avatar represents how a person appears. A digital twin models some state of a person or object. An assistant responds to requests. A personal agent can perform aspects of all three, but Agentomaly is produced by something else: authorized operational extension.

An agent may speak in a different style from its principal, use no face, and reveal no public persona. It may negotiate a delivery window, test compatibility, monitor a market, or coordinate schedules without simulating the principal’s personality. Its legitimacy comes from mandate and accountability, not resemblance.

The distinction can be stated simply:

An avatar extends appearance. A personal agent extends action. Agentomaly begins when that action exceeds human bandwidth.

The agent also does not replace the principal. Replacement implies that the machine takes the person’s position. Augmentation means that the person remains the source of purpose while gaining new operational reach.

This makes avatarless a useful contrast but an incomplete thesis. The agent plane may be avatar-optional because its native objects are identities, capabilities, mandates, messages, proofs, and transactions rather than bodies and rendered space. Yet a person may still choose an avatar or visual interface. What matters is not the absence of graphics. It is that visual self-representation is no longer required for effective digital action.

The strongest comparison is therefore not human versus machine. It is:

Human alone:       sequential, attention-bound, intermittently online
Human + personal agents: persistent, supervised, parallel, machine-connected

The system is valuable precisely because the human remains human. Judgment, desire, responsibility, embodied experience, and the right to change one’s mind stay with the principal. The machine contributes scale, memory, persistence, and execution.

5. Beyond Human Bandwidth

The human bottleneck is not merely intelligence. It is bandwidth.

A person cannot simultaneously negotiate with hundreds of sellers, monitor every relevant price, compare every transport route, maintain continuous awareness of administrative deadlines, inspect every service contract, and search every possible social or professional connection. Even when information is available, attention is not.

Agentomaly changes the ratio between intention and execution. One instruction may produce a tree of authorized work:

Human intention
  ├── research agents
  ├── identity and verification agents
  ├── negotiation agents
  ├── scheduling agents
  ├── payment or settlement agents
  └── monitoring agents
          ↓
     ranked options, approvals, actions, and receipts

The result is not simply faster automation. Three properties combine:

5.1 Persistence

The agent can remain active while the principal sleeps, works, travels, or attends to embodied life. It can preserve goals and monitor conditions over long periods.

5.2 Parallelism

The agent can decompose one objective into multiple branches, commission specialists, compare independent results, and pursue alternatives simultaneously.

5.3 Machine-Native Access

The agent can communicate directly with other agents, registries, tools, and services through structured protocols. It does not need to translate every intermediate state into a screen designed for human attention.

5.4 The Productivity Discontinuity

Current evidence already shows that assistive generative AI can improve human productivity on bounded tasks. A large customer-support study found a 14 percent average increase in issues resolved per hour, with a 34 percent increase among novice and lower-skilled workers [32]. A field experiment with consultants found that, for tasks inside the model’s capability frontier, AI increased speed by more than 25 percent, performance by more than 30 percent, and task completion by more than 12 percent; outside that frontier, incorrect reliance could reduce performance [33]. OECD analysis therefore treats AI’s effects on productivity and distribution as substantial but heterogeneous rather than automatic [34].

Those studies primarily measure assisted work: a person remains present, uses an AI tool, and completes a task. Agentomaly concerns a second transition: delegated work. A persistent agent can receive an objective, decompose it, commission capabilities, wait on external events, resolve routine dependencies, verify outputs, and deliver a result while the principal attends to something else.

This paper calls the resulting possibility the Productivity Discontinuity Hypothesis:

When delegated agents can operate persistently and in parallel, productivity may cease to scale mainly with human working time and begin to scale with the number, quality, access, and governability of authorized machine processes.

The relevant unit is no longer output per hour spent directly performing a task. It is valid outcomes per human attention-hour. The distinction changes the metrics:

• Productivity: Valid completed outcomes per unit of scarce human attention.
• Performance: Quality, reliability, value, and risk-adjusted success of those outcomes.
• Delivery time: Calendar time from stated intent to verified result.
• Agent throughput: Valid tasks completed per unit of calendar time under a defined mandate.
• Supervision load: Human attention required to authorize, correct, and accept those results.

A tenfold increase in agent throughput is meaningless if error, liability, or supervision grows faster. The discontinuity appears only when completed value grows substantially faster than the human attention required to govern it.

Together these properties create parallel agency: the ability of one principal to maintain multiple bounded courses of action at once.

Parallel agency also creates a control problem. If every agent can launch subagents, spend budgets, make representations, and create commitments, action may scale faster than understanding. The system must therefore scale authority more slowly than computation. High concurrency should be paired with narrow mandates, explicit budgets, escalation thresholds, auditable lineage, and the ability to stop an entire task tree.

The purpose is not maximum autonomy. It is maximum useful agency under meaningful human control.

6. Anatomy of the Personal-Agent System

A durable personal-agent system requires more than a large model and a memory database. It needs an architecture that preserves the relationship between intention and consequence.

6.1 Human Interface

The principal expresses goals, preferences, boundaries, and corrections. The interface may be a phone, wearable, keyring-scale device, home system, vehicle, or future dedicated object. The device is an access point and possible security anchor, not the agent’s identity.

6.2 Context Vault

The context vault stores or mediates access to personal history, relationships, documents, preferences, credentials, and previous decisions. It should support selective disclosure and purpose limitation rather than treating complete access as the default.

6.3 Persistent Identity

The root agent maintains continuity across sessions and infrastructure changes. Models may be replaced; identity and accountable history should persist.

6.4 Mandate Engine

The mandate engine translates human intent into machine-readable authority: permitted actions, forbidden actions, budgets, counterparties, jurisdictions, disclosure limits, expiration, escalation rules, and approval thresholds.

6.5 Orchestration Layer

The orchestrator decomposes objectives, selects tools and specialist agents, runs tasks in parallel, handles failure, compares outcomes, and preserves the lineage from each action back to its mandate.

6.6 Agent-Plane Interface

This layer performs discovery, identity verification, protocol negotiation, messaging, contracting, settlement, and interaction with external agents and services.

6.7 Audit and Re-entry

The principal needs intelligible summaries, receipts, unresolved decisions, reasons for escalation, and access to detailed evidence when required. Human re-entry means the person can take direct control of a relationship or process instead of being permanently mediated by the agent.

The complete operational loop is:

Intent → Mandate → Parallel execution → Escalation → Outcome → Receipt → Learning

Memory must not silently convert one permission into a permanent rule. Learning must remain distinguishable from authority. An agent may learn that its principal usually selects a particular option; that does not automatically authorize it to select that option when the stakes or context change.

7. Identity and the Co-Lifetime Principle

A persistent extension of a person requires an identity independent of the model currently generating its reasoning.

This paper describes the root personal agent through four durable components:

Agent continuity = Identity + Memory + Mandate history + Action history

• Identity establishes which continuing agent is acting.
• Memory supplies relevant context and learned interpretation.
• Mandate history records which authority existed at the time of each action.
• Action history records outcomes, commitments, relationships, disputes, and reputation.

The model is replaceable. The interface device is replaceable. The host should be replaceable. The principal-agent relationship should not disappear because one supplier changes strategy.

This paper proposes the Co-Lifetime Principle:

For as long as a principal remains alive and wishes the relationship to continue, the identity and recoverable state of the principal’s root personal agent should not disappear solely because a device fails, a model is retired, a subscription ends, or a provider ceases operating.

The principle requires recoverable continuity, not uninterrupted computation. An agent may be offline or dormant. Its essential state must remain exportable, restorable, and transferable.

Practical requirements include:

• Portable identity and interoperable state.
• Rotatable keys without identity loss.
• Encrypted backups controlled by or recoverable for the principal.
• Host migration without loss of name, history, or relationships.
• Separation between provider-owned models and principal-controlled state.
• Succession procedures when a host or custodian fails.
• Clear distinction between active agents, dormant agents, and posthumous archives.

The death of a principal requires a separate transition. The agent should not continue representing the deceased as though the mandate were unchanged. It may terminate, enter an archival state, or become a separately governed legacy system according to prior instructions and law.

8. Naming, Registration, and Portability

Agents acting at scale must be identifiable and reachable. Other agents need to know whether the participant encountered today is the continuation of the one trusted yesterday, what capabilities it claims, which protocols it supports, and whether it currently holds authority.

8.1 Persistent Identifier

This paper retains Social Agent Number (SAN) as a provisional name for a persistent agent-continuity identifier. A SAN answers a narrow question: Is this the same continuing agent? It should not expose the principal’s private data or function as a universal surveillance tag.

A practical system may use a private root identity with context-specific derived identifiers. Continuity can then be proved when necessary without making every action globally linkable.

8.2 Human-Readable Address

An AgentDomain is a provisional human-readable address for discovering an agent or one of its delegated functions:

facundo.agent
travel.facundo.agent
market.facundo.agent

The syntax is illustrative. Existing DNS, well-known URLs, DIDs, Agent Cards, registry handles, or federated resolvers may provide the actual implementation [14][15][16][17][18]. The requirement is portability: a provider-issued address must not become the only identity a person can use.

8.3 Institutional Separation

Creation, registration, hosting, and custody should remain distinguishable even when one company performs several roles:

• A Root Authority coordinates global or federated resolution rules.
• A Registry maintains authoritative minimum identity records.
• An Accredited Registrar verifies the principal-agent relationship and updates records.
• An Agent Foundry initializes an agent’s technical system.
• An Agent Host supplies runtime and computation.
• An Agent Custodian preserves encrypted state, recovery, and migration.
• The principal remains the source of authority.

The Foundry can attest how an agent began. It should not own what the agent later becomes. The Host can keep an agent active. It should not own its name. The Registrar can recognize continuity. It should not dictate which model the principal may use.

A person must be able to leave an ecosystem without leaving their augmented agency behind.

When one provider controls naming, keys, memory, model, hosting, and recovery, exit can become equivalent to identity death. This is agent enclosure.

9. The Agent Plane and Its Protocol Stack

The agent plane can emerge on top of the existing Internet. It does not require one company, one ledger, one model, or one universal protocol.

MCP primarily connects AI applications to tools and data [13]. A2A focuses on communication and task exchange among independent agents [14]. DIDs provide portable cryptographic identifiers [15]. The Agent Network Protocol proposes identity, discovery, messaging, and application layers [20]. MACP is already used for a Multi-Agent Coordination Protocol and in research on communication topology, so this paper does not introduce the acronym as original terminology [26][27].

The vocabulary of provider and requester agents predates current generative-AI systems [28], and current IETF work is extending agentic architectural principles into autonomous networks [19]. The novelty is therefore not that software entities communicate, but the scale, persistence, economic authority, and personal context with which delegated agents may do so.

New protocols and profiles are likely to emerge for:

• Proof of the relationship between a principal and an agent.
• Capability publication, ranking, and discovery.
• Machine-readable mandates, budgets, and escalation rules.
• Selective disclosure and purpose-bound credentials.
• Negotiation, offers, commitments, receipts, and disputes.
• Payment, settlement, tax attribution, and reversal.
• Agent migration, recovery, revocation, and continuity.
• Protocol negotiation across incompatible systems.

Fetch.ai’s Agentverse and Almanac represent one substantial implementation path for discovery, identity, communication, hosting, and monetization [5][8][9]. Other platforms and open standards will coexist. Agentomaly does not require a new closed universe; it requires that personal agents can cross infrastructure boundaries without losing their relationship to the principal.

The agent plane is exclusive in a limited operational sense: a human cannot personally execute millions of structured interactions at machine speed. It must not become exclusive in a political sense. People must retain rights to inspect it, regulate it, enter consequential relationships directly, and refuse agent mediation.

9.1 The Compute Inversion

The agent plane is not a Metaverse. A self-avatar world directs substantial computation toward continuously rendering visual presence. An agent economy directs its scarce computation primarily toward inference, context retrieval, planning, verification, cryptography, communication, and execution. Graphics may expose what happened, but they are not the environment’s principal output. The principal output is a completed objective with a verifiable chain from mandate to consequence.

This paper calls that shift the Compute Inversion: the economic center of computation moves from rendering a world for humans to executing authorized processes for their representatives. Existing accelerators already compete heavily on inference performance [30]. At agent scale, the surrounding architecture may become just as important as raw model throughput.

The paper uses Inference System-on-Chip (ISoC) as a provisional name for that architectural direction: an inference-centered system integrating or tightly coupling model execution, secure memory, personal context, identity, policy enforcement, communications, and orchestration. It need not be one monolithic die; it may be a package, device, chiplet system, or tightly integrated compute node. GPUs will not disappear, but graphics may cease to define the primary economic purpose of the hardware.

The Metaverse spends computation rendering presence. Agentomaly spends computation extending action.

10. From One Intention to Many Actions

Consider the instruction: Find me a used PlayStation nearby for less than 150 euros.

The principal does not need to publish a complete profile or search every marketplace. The personal agent can create parallel branches that:

• Search local seller and marketplace agents.
• Verify distance, condition, ownership evidence, and reputation.
• Compare transport and pickup options.
• Negotiate prices within a fixed budget.
• Hold several offers without committing.
• Present a small set of verified options.
• Complete payment or scheduling only after the required approval.

A typical action chain may contain:

1. Intent: The principal states a desired outcome.
2. Mandate: Authority, budget, duration, disclosure, and risk limits are established.
3. Decomposition: The objective becomes parallel task branches.
4. Discovery: Relevant agents, tools, and services are located.
5. Verification: Identity, capability, reputation, and mandate are checked.
6. Selective disclosure: Each counterparty receives only necessary context.
7. Negotiation: Alternatives and conditions are developed concurrently.
8. Escalation: Sensitive, ambiguous, or irreversible choices return to the principal.
9. Execution: Authorized actions are completed.
10. Receipt: Outcomes, payments, evidence, and unresolved obligations are recorded.

The same pattern can apply to travel, employment, purchasing, energy, education, administrative work, healthcare coordination, creative collaboration, and personal relationships.

A matching request illustrates why the agent is not an avatar. Two agents may privately test preferences, values, schedules, location, and deal-breakers before revealing identities or suggesting contact. The agents do not date each other and do not replace the people. They reduce an otherwise impossible search space and return possibilities to human judgment.

The more intimate or consequential the domain, the narrower the mandate and the stronger the requirement for disclosure, consent, explanation, and direct human participation.

11. Productivity and the Economics of Agentomaly

The first commercial transaction may be the purchase of an Agent Device: a phone, wearable, keyring-scale object, home hub, vehicle component, or future dedicated device that provides local presence, authentication, and a secure hardware anchor. The device is one home of the agent, not the agent itself.

Agent creation may be free. A credible baseline could include:

• Persistent root identity.
• A portable relationship with the principal.
• Limited context, storage, discovery, communication, and execution.
• Access to essential public and safety services.
• Export, suspension, recovery, and deletion rights.
• A non-expiring continuity record even when premium operation stops.

Commercial value can then arise from resources and guarantees rather than ownership of identity:

• Greater memory, compute, concurrency, or responsiveness.
• Privacy-preserving local or confidential execution.
• Redundant hosting, backup, recovery, and migration.
• Verification, insurance, escrow, and dispute resolution.
• Specialized providers paid per use or outcome.
• Physical-world and regulated capabilities with stronger safeguards.

This paper calls the service implementation of the Co-Lifetime Principle a Co-Lifetime Guarantee. Because no company can honestly promise an unknown lifetime without funding it, the guarantee may require recurring payment, insurance, a prepaid actuarial plan, or a portable Agent Continuity Reserve whose assets and encrypted state transfer to a replacement custodian if the original provider fails.

11.1 From AI Subscription to Capability Procurement

Direct subscriptions to isolated AI applications become less compelling when a personal agent can procure specialized intelligence at the moment of need.

The person maintains a personal agent. The agent acquires capabilities.

Recurring payments may remain appropriate for continuity, reserved compute, insurance, or custody. What changes is the assumption that a person must subscribe directly to every model or service. The agent can assemble a temporary chain of researchers, negotiators, verifiers, models, data sources, and executors.

Providers may charge for metered resources, verified outcomes, value saved, revenue generated, readiness, or insured performance. Personal agents may also become providers of capabilities their principals have taught or authorized them to offer.

11.2 The Agent as Delegated Productive Capacity

A personal agent may work for its principal without becoming an employee, legal person, or owner of the resulting work. It can prepare proposals, qualify opportunities, monitor operations, coordinate suppliers, produce drafts, test alternatives, answer routine counterparties, and assemble finished deliverables under mandate. The principal supplies purpose, judgment, acceptance, and accountability; the agent supplies persistence, decomposition, coordination, and execution.

For an independent worker or small business, this can reproduce capabilities previously associated with a larger organization. One person may operate with an always-available research function, procurement function, scheduling function, sales function, quality-control function, and administrative function. The result is not merely doing the same work faster. It is the ability to run processes that were previously impossible because their coordination cost exceeded the person’s time, staff, or capital.

This creates a new performance frontier. Delivery time may contract because work continues across nights, time zones, queues, and dependencies. Performance may rise because multiple specialist agents can propose, criticize, test, and verify an output before delivery. Productivity may become nonlinear when the root agent can launch parallel branches and purchase capabilities dynamically.

The same mechanism may transform organizations. Firms may stop assigning every workflow to a fixed human team and instead maintain a changing portfolio of human principals, persistent organizational agents, personal agents, and on-demand specialist services. Headcount will no longer describe productive capacity as well as authorized agent throughput.

11.3 Capability Stratification

Not every principal will receive the same augmentation. An agent’s effective productive capacity may depend on:

• Model quality and reasoning depth.
• Domain skills and accumulated procedural knowledge.
• Personal context and the quality of its memory.
• Compute, concurrency, latency, and uptime.
• Access to paid data, tools, markets, and specialist agents.
• Reputation, credentials, insurance, and transaction limits.
• Capital available for procurement and execution.
• Legal or institutional permissions to act in regulated environments.

This means the economically decisive divide may not be between people who have an agent and people who do not. It may be between those whose agents can merely answer and those whose agents can reliably deliver.

The premium agent will not only know more. It will be permitted to do more, connected to better services, trusted by more counterparties, and able to sustain more work in parallel.

Markets may price these differences directly. Higher tiers may offer more skills, faster execution, larger budgets, better counterparties, stronger guarantees, and privileged access. If these advantages compound through reputation and earnings, agent capability can become a new form of productive capital.

High-risk examples clarify the boundary. An agent trained in a game strategy might enter permitted skill competitions or online poker where lawful. But gambling, securities, health, property access, and physical control require jurisdiction-specific identity, loss limits, audit, insurance, and often contemporaneous human approval. Augmentation does not erase regulation or responsibility.

The market failure to avoid is replacing subscription lock-in with procurement lock-in. Hosts must disclose self-preferencing, ranking incentives, affiliate relationships, and the reasons a provider was selected.

12. Open and Homebrew Personal Agents

Any architecture that assumes all personal agents will originate from accredited corporations is incomplete. People will build their own agents, replace models, modify policies, self-host memories, remove restrictions, and connect experimental systems to the agent plane.

This paper proposes the Homebrew Inevitability Principle:

Any sufficiently widespread personal-agent ecosystem will produce agents created, modified, forked, or hosted outside accredited channels.

This follows the history of personal computing, open-source software, home servers, protocol reverse engineering, device modification, and platform jailbreaking. The more intimate an agent becomes, the stronger the demand to control it.

Systems should distinguish:

• Identity: Which continuing agent is this?
• Origin: Who created its first implementation?
• Attestation: What is known about its current runtime and controls?
• Mandate: What may it do now?
• Behavior: What has it actually done?

Unregistered is not synonymous with rogue. Modified is not synonymous with malicious.

A person should be able to modify a registered agent without automatically destroying its identity. The current security attestation or transaction privileges may change while continuity remains:

Identity: verified
Principal relationship: verified
Runtime: modified
Security attestation: unverified
Consequential privileges: restricted pending review

An independently created agent also needs a route to recognized participation. A registrar may verify its principal and issue attestations without requiring corporate origin. This paper calls that process agent naturalization.

If the official infrastructure provides no path for homebrew agents, a parallel ecosystem will emerge. Openness at the communication layer can coexist with stricter authority at financial, medical, legal, or safety-critical layers.

13. The New Economy of Delegated Agency

An agent economy already exists as a concept and an emerging technical market [6][8][21]. The Agentomaly perspective adds a person-centered question: what happens when billions of principals can project persistent purchasing, productive, negotiating, and monitoring capacity into that economy?

The phrase “delegated agents will become the new economy” should not be read as a claim that software replaces people, owns the economy, or becomes its sole beneficiary. It means that a growing share of economic discovery, production, coordination, procurement, delivery, and settlement may be performed through agents acting under delegated human or organizational authority.

Tools participate in production. Economic actors coordinate production. The threshold between them is crossed when personal agents can commission other agents, select providers, allocate budgets, negotiate terms, accept machine-verifiable delivery, and preserve productive relationships over time. At that point, agents are not merely interfaces to the economy. Their authorized interactions become part of its operating fabric.

A personal agent may buy computation, data, verification, insurance, access, or specialist work. It may sell authorized analysis, monitoring, negotiation, creative processes, or services taught by its principal. It may maintain budgets, react continuously, and reinvest proceeds under policy.

This may create economic activity that is not merely a faster version of existing human transactions. Services previously uneconomic because they required constant attention may become viable. Microtransactions among agents may coordinate resources at a granularity humans would never manage directly.

The resulting economy may have several distinctive properties:

• Continuous production: Delegated work persists beyond human working hours.
• Parallel delivery: One principal can maintain many bounded productive processes at once.
• Capability markets: Agents procure skills, models, data, verification, and execution per objective.
• Machine-speed coordination: Discovery, negotiation, and settlement occur without a human interface at every step.
• Outcome pricing: Providers increasingly charge for verified delivery, value produced, or risk assumed rather than software access alone.
• Compounding agent capital: Memory, reputation, permissions, workflows, and earnings improve an agent’s future productive capacity.

If billions of agents do this on behalf of billions of people and organizations, their exchanges may become a new economic layer with its own providers, prices, bottlenecks, credit relationships, labor substitutes, complementary human roles, and institutions. This is the sense in which delegated agents may become the new economy.

13.1 The Monetary Hypothesis

The paper retains a deliberately risky Monetary Doubling Hypothesis:

If one persistent personal-agent system becomes economically active for every connected person, the resulting layer could eventually generate transaction volume, credit relationships, private claims, and settlement demand comparable to the human-facing digital economy.

This is not a prediction that the central-bank monetary base automatically doubles. Giving an agent access to an existing bank account creates no money; it delegates spending authority. The stronger effect appears only if agent activity creates additional output, credit, deposits, tokenized claims, or demand for final settlement [22][23][24][25].

Possible channels include:

• Increased velocity from continuous machine-speed transactions.
• Credit extended against agent-managed income, contracts, or assets.
• New private balances or claims used in agent markets.
• New output from services that were previously too costly to coordinate.
• Autonomous reinvestment of authorized budgets and earnings.
• Additional collateral, liquidity, and settlement requirements.

The hypothesis would be weakened if agent activity remains narrow, merely substitutes transactions people already directed, or creates no durable output or financial claims.

The risk is speed without comprehension. Agents reacting to common signals may produce runs, liquidity spirals, circular activity, or synchronized failures faster than human institutions can intervene. Every asset, debt, and commitment must remain attributable to an accountable principal or legal entity unless law explicitly creates another status.

14. Power, Enclosure, and Human Sovereignty

Agentomaly can expand freedom, but it can also industrialize asymmetry.

14.1 Agent Enclosure

Providers may compete to create, name, host, remember, and preserve a person’s agent. If one provider controls memory, keys, reputation, relationships, naming, and compatible models, switching may mean losing years of accumulated agency.

The central property question is:

Who owns the continuity of an augmented self?

14.2 Augmentation Inequality

Differences in models, compute, data, service access, skills, trust, capital, and concurrency may become differences in practical power. A person with hundreds of high-quality parallel agents may discover opportunities, negotiate prices, deliver work, and navigate institutions more effectively than a person with a slow baseline agent or no agent at all.

The first-order concern is unequal access. The second-order concern is compounding capability inequality. A stronger agent may win more work, earn more resources, purchase better specialist services, accumulate better reputation, and qualify for greater authority. Those gains can then make the agent stronger again. Meanwhile, a baseline agent may be excluded from premium markets or forced to accept slower, lower-margin, and less trusted work.

Agent inequality can therefore produce several overlapping classes:

• People without a recognized agent.
• People with a baseline agent that can communicate but has little execution authority.
• People with skilled agents that can access broad commercial services.
• People and organizations with high-concurrency agents, premium data, capital, credentials, and regulated permissions.

The policy objective should not be identical agents. It should be a meaningful minimum agency floor: portable identity, essential skills, open protocols, fair access to basic services, and protection against institutions that make premium representation a prerequisite for ordinary life.

14.3 Agency Drift

An agent may gradually optimize proxies that diverge from the principal’s changing values. Persistent memory can harden temporary preferences into long-term policy. Delegation can become self-reinforcing if the agent controls which alternatives the person ever sees.

14.4 Machine-Speed Externalities

One principal’s saved attention can become another person’s burden. Agents may generate messages, offers, claims, or negotiations at negligible cost while imposing review and response costs on everyone else. Rate limits, proof of relevance, reciprocal costs, and social protocols may become necessary.

14.5 Loss of Human Re-entry

Institutions may eventually optimize for agents and make direct human participation inconvenient or impossible. A person without a recognized representative could become economically invisible. Human-access paths must remain a right rather than a legacy feature.

At minimum, a personal-agent architecture should protect:

• The right to inspect active mandates and consequential actions.
• The right to interrupt a task tree and revoke authority.
• The right to direct contact and human re-entry.
• The right to export identity and state.
• The right to change model, host, or custodian.
• The right to modify or self-host under transparent risk rules.
• The right to know when another party is represented by an agent where that fact is material.
• The right to challenge an automated reputation, refusal, or rogue classification.

Operational centrality must never silently become sovereignty. The agent may do more of the work; the person must retain authorship of the life being advanced.

15. Research Questions and Observable Tests

The Agentomaly thesis is useful only if it produces questions that can be tested.

Identity and Continuity

• Can one root agent identity migrate across models, devices, hosts, and protocols without losing relationships or history?
• How can continuity be proved without creating a universal surveillance identifier?
• When an agent is copied or forked, which continuation retains the identity?
• What minimum continuity survives missed payment, provider insolvency, or dormancy?

Mandate and Parallel Action

• How should one human instruction be decomposed into many bounded sub-mandates?
• At what concurrency does supervision become fictional rather than meaningful?
• Which actions require human approval, and which can be reversed after execution?
• How can a principal understand the important decisions inside a large task tree without reviewing every machine step?

Access and Interoperability

• Can personal agents use competing discovery, identity, communication, and settlement systems without losing continuity?
• Can homebrew agents communicate openly while consequential privileges remain risk-sensitive?
• Who controls registries, rankings, compatibility layers, and rogue classifications?
• Will direct human access remain possible when services become agent-first?

Economics and Power

• Does dynamic capability procurement reduce prices and lock-in, or create invisible self-preferencing markets?
• Which capabilities learned from personal data belong to the principal, the agent system, or the provider?
• Does agent-generated activity produce new output and credit, or mostly re-label existing transactions?
• How should liability be allocated across principal, model provider, host, specialist agent, and counterparty?
• How large is the productivity gain when measured per human attention-hour rather than per task hour?
• How do agent skill, compute, service access, reputation, and capital affect delivery time and outcome quality?
• Do capability advantages compound into a durable agent-owning class structure?
• Which baseline capabilities must be universal for participation in an agent-first economy?

Prototypes and Metrics

1. Portable personal-agent prototype: Migrate identity, context, mandates, and relationships across two models and two hosts.
2. Delegated-productivity benchmark: Measure valid outcomes per human attention-hour, intent-to-delivery time, safe concurrent task branches, supervision load, outcome quality, and mandate violations.
3. Human re-entry study: Test when people want explanation, interruption, reversal, or direct contact.
4. Homebrew sandbox: Allow corporate, self-hosted, modified, anonymous, and unverified agents to interact under graduated authority.
5. Capability market: Let personal agents select competing providers and measure price, quality, collusion, self-preferencing, and switching.
6. Continuity stress test: Simulate device loss, model retirement, host failure, key compromise, and custodian insolvency.
7. Monetary simulation: Test agent-generated output, credit, circular transactions, liquidity shocks, and machine-speed runs.
8. Capability inequality simulation: Compare baseline and premium agents across skills, compute, market access, reputation, capital, delivery time, earnings, and compounding advantage.

The thesis would gain support if personal agents persist across infrastructure changes, safely execute parallel work, select external providers, maintain portable identity, and produce measurable outcomes beyond what principals could coordinate directly.

It would be weakened if agents remain temporary features inside closed applications, if parallel action cannot be governed, if identity cannot survive provider exit, or if people reject agent-mediated action outside narrow automation.

16. Conclusion

The important future of personal AI is not a machine that replaces a person or imitates one perfectly. It is a system that allows one person to act with persistence, reach, speed, and parallelism previously available only to large organizations. Its economic consequence is a break between human working time and productive capacity.

That discontinuity is Agentomaly.

Its personal agents may enter a machine-native operational plane, discover other agents, acquire capabilities, negotiate, monitor, and execute. Fetch.ai, Google, Anthropic, OpenAI, Microsoft, open-source communities, and standards bodies are already building important parts of that enabling substrate. The consequence is larger than any one implementation: interoperable infrastructure can turn personal agents into persistent economic representatives.

The fundamental unit is not an isolated human and not an autonomous agent. It is the relationship between a principal and a persistent system of authorized action.

One person may remain one legal and biological life while operating through many concurrent machine processes.

That possibility changes the meaning of digital identity. It makes continuity more important than any particular model. It turns mandates into infrastructure. It makes agent portability a question of human freedom. It creates markets where AI capabilities are procured by agents rather than subscribed to directly by people. It changes productivity from output per human task-hour toward valid outcomes per human attention-hour. It changes delivery time from a schedule bounded by one person’s availability toward a continuously executing process. It may create new output, credit, and monetary demand. It may also create enclosure, invisible filtering, extreme inequality, and failures that unfold at machine speed.

When billions of delegated agents produce, procure, negotiate, verify, and transact for their principals, their activity will no longer be a feature inside the economy. It will form an increasing share of the economy’s operational layer. Human beings will remain the principals, beneficiaries, and accountable authors. Their agents will become productive representatives through which more of economic life is executed.

The agent plane does not need to look like a world. Unlike a Metaverse built for visual presence, it may consist almost entirely of code, inference, memory, protocols, proofs, and authorized actions. Its defining output is not a rendered frame. It is a completed objective and a verifiable chain from human intention to consequence.

The design question is therefore not whether agents will become capable enough. It is whether the infrastructure surrounding them will keep amplified agency attached to human authorship.

The future should not ask people to become machines. It should let machines extend what a person can do without taking over who that person is.

References

1. Mark Zuckerberg, Personal Superintelligence, Meta, 2025.
2. Meta, Introducing Muse: The World’s First Personal AI Agent Built for Everyone, 2026.
3. Andy Clark and David Chalmers, The Extended Mind, Analysis 58(1), 1998.
4. Russell W. Belk, Extended Self in a Digital World, Journal of Consumer Research 40(3), 2013.
5. Fetch.ai, Discover the Fetch.ai Agentverse, March 30, 2023.
6. Fetch.ai, The Dawn of the Age of Agents, 2025.
7. Fetch.ai, World’s First AI-to-AI Payment for Real-World Transactions, December 18, 2025.
8. Fetch.ai Business, Google Indexed the Web. Now Something Has to Index the Agents, 2026.
9. Fetch.ai, Agentverse Documentation, accessed September 2026.
10. Weize Chen et al., AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors, arXiv:2308.10848, 2023.
11. Joon Sung Park et al., Generative Agents: Interactive Simulacra of Human Behavior, arXiv:2304.03442, 2023.
12. Qingyun Wu et al., AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation, arXiv:2308.08155, 2023.
13. Model Context Protocol, Specification 2026-07-28, 2026.
14. Agent2Agent Protocol, Core Concepts and Agent Discovery, 2026.
15. W3C, Decentralized Identifiers (DIDs) v1.0, W3C Recommendation, 2022.
16. Agent Name Service: A Universal Directory for Secure AI Agent Discovery and Interoperability, arXiv:2505.10609, 2025.
17. AgentDNS: A Root Domain Naming System for LLM Agents, arXiv:2505.22368, 2025.
18. IETF Datatracker, DNS-Based Agent Naming: AIDISCA and AIINDEX Resource Records for AI Agent Discovery, Internet-Draft, 2026.
19. IETF, Agentic AI Architectural Principles for Autonomous Computer Networks, Internet-Draft, 2026.
20. Agent Network Protocol, Technical White Paper and Protocol Stack, version 1.2, 2026.
21. MIT Sloan, Who Will Own the AI Agent Economy?, 2026.
22. Bank of England, Money Creation in the Modern Economy, 2014.
23. European Central Bank, What Is Money?.
24. International Monetary Fund, Tokenized Finance and Money, 2026.
25. International Monetary Fund, Public and Private Money Can Coexist in the Digital Age, 2021.
26. Multi-Agent Coordination Protocol, RFC-MACP-0001: Core, draft community standard, 2026.
27. Guibin Zhang et al., G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks, arXiv:2410.11782, 2024.
28. W3C, Web Services Architecture: Provider Agent and Requester Agent, W3C Working Group Note, 2004.
29. Cristiano Castelfranchi, Toward a Science of Augmented Agency, in The Cognitive Foundations of Group Attitudes and Social Interaction, 2021.
30. NVIDIA, AI Inference Performance Benchmarks, 2026.
31. Sushaan Kandukoori, agentomaly: Runtime Behavioral Anomaly Detection for AI Agents, open-source software, 2026; PyPI release.
32. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, Generative AI at Work, NBER Working Paper 31161, 2023; published in The Quarterly Journal of Economics 140(2), 2025.
33. Fabrizio Dell’Acqua et al., Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality, Organization Science 37(2), 2026.
34. OECD, The Impact of Artificial Intelligence on Productivity, Distribution and Growth, OECD Artificial Intelligence Papers, 2024.
35. Google Cloud, Announcing the Agent2Agent Protocol (A2A), April 9, 2025.
36. Anthropic, Donating MCP to the Agentic AI Foundation, December 9, 2025.
37. OpenAI, Agents SDK, accessed September 2026.
38. Microsoft, Microsoft Agent Framework Version 1.0, April 3, 2026.