A short speculative novel, followed by an economic stress test. Horizon: 2028–2030. Evidence checked: 14 September 2026. About 18–20 minutes including the interactive reader.
People could own a portable system that turns their experience into useful services. Whether they earn lasting income from it depends on the cost of replacing them, the quality of their evidence, and their power to refuse a bad contract.
The question
What would a world look like if each person could grow a distinctive knowledge system, give it an agent, and license that agent to organizations and robots?
Here, AES means an agentic engineering system: agents, tools, knowledge, verification, and accountable people working together. Extending it beyond software into physical operations is a scenario assumption. Knowledge stock means accumulated usable understanding, not organization shares. “Own” means control of records, permissions, exports, and licensed services; it does not assume ownership of every underlying model or every fact learned at work.
Reading key: the characters, dialogue, contracts, payments, and events below are fiction. The evidence section reports existing sources. Everything labeled Codex inference is a proposed design, incentive model, or forecast. Game theory establishes consequences of assumptions; it cannot prove that this future will happen.
Background
The starting idea is simple: keep knowledge in portable records that survive a change of application. Search indexes can be rebuilt. An agent can change without replacing the owner's history. Useful knowledge also needs intent, context, boundaries, and evidence of what happened when it was used.
The new connection is economic. A individual record becomes a service when someone else can request a bounded result, assess its quality, and pay for authorized use. A second connection is physical: a robot may possess the movement skills while a person supplies knowledge of a particular place, failure, or exception.
Fiction · 2028 · The first fee
The first time Mai's agent earned money, it told a machine to do nothing.
The repair robot had arrived before sunrise. It stood beside a pump in a riverside workshop, one wheel on a dry patch of concrete. Its camera could read the model number. Its service software knew the maintenance manual. The replacement part was correct.
Mai watched the inspection feed from her kitchen.
“Ask it to show the paint below the housing,” she said.
Her agent, Thread, requested another photograph. Under the blue paint was a crooked line of white. Someone had repainted the unit after the last flood. The manual described the factory machine; it did not describe what this machine had become.
Thread sent a recommendation: stop the planned work and request a qualified site inspection. It attached the image, the discrepancy, and the limits of its conclusion. The robot's local controller kept it still. The site operator made the decision.
Mai received a small fee for a verified exception report. She had not sold the workshop her notebooks. She had licensed one assessment.
Two years earlier, Thread had been little more than a search box over her files. It remembered invoice numbers better than she did and misunderstood the reasons for half her decisions. Mai had once written “same as last time.” Thread had treated that as knowledge. She had learned to write what was the same, what had changed, and how she knew.
Now, after a repair, it asked three questions. What surprised you? What did you notice before the instruments did? What would make your explanation wrong?
Mai answered in her own language. Thread preserved the recording beside its summary. A claim remained provisional until a later observation supported it. Two contradictory cases stayed contradictory. It did not turn uncertainty into agreement just to make the notebook tidy.
The valuable part was growing slowly: not a collection of pump manuals, but a history of when the manuals had been insufficient.
Her son asked whether Thread was becoming her.
“It is learning what I have managed to explain,” she said. “That is smaller.”
Before she went to work, she opened its first receipt. The fee was modest. Below it was a more interesting line: Assessment accepted. Operator requested availability next month.
Someone wanted access to her future judgment.
Fiction · 2029 · The copy clause
By the following year, Thread served three maintenance firms. One used it through an agent directory. Another sent structured cases directly. A third still required a human to telephone Mai whenever the agent declined a task.
There was no universal agent labor market. There were directories that disagreed about credentials, invoices that arrived late, and customers who thought “autonomous” meant “someone else's responsibility.”
The largest firm offered a better rate. Thread highlighted a clause: the customer could retain all outputs and use them to train a replacement service.
“They want a discount on replacing you,” Mai said.
The procurement manager was less theatrical. His firm could not depend forever on one supplier. Equipment needed support when people became ill, changed careers, or disagreed over price. He wanted continuity. She wanted an income that survived the next model release. Both positions made sense.
Mai refused the broad clause. She offered a narrower arrangement: case records could be retained for maintenance and audit; training rights required a separate purchase; routine procedures could be delivered as a versioned offline package; unusual cases would still require live consultation. Critical operations would have a fallback supplier.
It cost her the contract.
For six weeks, the platform's cheaper service answered the firm's familiar questions adequately. Mai had overestimated how much of her work was rare. Her income fell. She stopped describing every note as an asset.
Then a new batch of parts began producing failures that did not resemble the old cases. Her notebooks did not contain the answer either. But she knew which measurements were missing, whom to ask for a second opinion, and which early explanations were weak.
The firm returned for a limited investigation. It did not buy access because Mai possessed an uncopiable past. It bought access because she was still learning.
She accepted a lower fee than she first wanted. The firm accepted a narrower license than it first wanted. Both had alternatives; neither had a monopoly.
That evening, Mai removed an old instruction from Thread: Prefer an answer when confidence is reasonable.
She replaced it with: Explain what evidence would change the decision. Decline when that evidence is missing.
Fiction · 2030 · A room with many owners
The cooperative occupied a former training room above a machine shop. Its most expensive object was a calibration rig. Its most argued-about object was a spreadsheet.
Twelve specialists had brought their agents. A buyer could contract with the cooperative once and receive several kinds of expertise. Members kept their own records. The cooperative maintained shared tests, handled disputes, and paid for evaluation equipment no member could easily afford alone.
The spreadsheet divided receipts between service providers, shared infrastructure, and an apprenticeship fund. It was a negotiated rule, revised by vote. Nobody pretended that a formula could perfectly identify whose idea had saved a machine.
An apprentice named An had little history to rent. If the market paid only for already proven expertise, she would always enter last. The cooperative therefore paid for supervised investigations and credited her new cases to her own system. Established members accepted the cost because they needed future colleagues and access to unfamiliar work. Some complained anyway.
Downstairs, a robot inspected a fixture. Its motion controller came from the manufacturer. Its work order came from the customer's AES. A materials assessment came from one member's agent; an exception check came from Mai's. The customer's supervisor authorized the next action.
The robot was not wearing a person's mind. It was using several limited services, each with a different boundary.
When the network failed, it followed its approved local fallback. The small cached procedure remained usable until its license expired. Mai could stop future requests; she could not make a delivered instruction disappear from a disconnected computer. The cooperative had learned to distinguish access control from magic.
At lunch, An asked why every observation was not for sale.
Mai showed her two folders. One contained maintenance cases. The other held family recipes, unfinished questions, and recordings of her mother telling stories. Thread could help her use both. Its commercial identity could see only the first.
“A useful life is larger than a useful service,” Mai said.
Outside, the city was uneven. Some people rented agents. Some were employees whose knowledge stayed inside organization systems. Some had uploaded years of experience and earned almost nothing. A few platforms collected fees from almost every exchange.
The cooperative had not solved that world. It had made one small part of it negotiable.
Before closing, Mai reviewed the day's new cases. An had found something she had missed. Mai asked permission to link the explanation into her own knowledge system, with An's name and its usage conditions attached.
For the first time, the older woman's agent paid the younger woman's agent to learn.
Codex inference · How a person builds this system
- 01 · ExperienceObserve something real
- 02 · MemoryKeep source + context
- 03 · JudgmentTest claims + exceptions
- 04 · individual agentApply scope + permissions
- 05 · AES / robotVerify before acting
- 06 · RenewalReturn permitted feedback
Records stay with the owner. The customer receives an authorized result. Renewal feeds the next experience.
The system begins as a useful memory for its owner. Commercial services are a later, selective layer. Building it only to sell knowledge would distort what gets remembered.
| Layer | What the person builds | What makes it useful |
|---|---|---|
| Experience | Consented notes, decisions, observations, demonstrations, and outcome records | The original event stays distinguishable from an AI summary |
| Understanding | Claims linked to sources, conditions, counterexamples, confidence, and unanswered questions | The system can explain when an idea applies and when it fails |
| individual fit | Goals, language, preferences, constraints, and things the agent must not decide | The system serves this person without assuming their preferences are universal |
| Demonstrated skill | Cases and procedures tested against held-out problems and a generic-agent baseline | A buyer can assess incremental value instead of counting documents |
| Delegation | A narrow agent with defined inputs, tools, outputs, spending limits, and escalation rules | A useful service can operate without exposing the owner's whole life |
| Renewal | Feedback, corrections, new cases, expiry checks, and independent review | The knowledge remains useful as the world changes |
The daily loop is experience → capture → question → test → revise → use → observe again. For physical skills, text may preserve the decision but miss timing, force, sound, and motion. Demonstrations and sensor records can help; they still require validation on the target robot and task.
Distinctiveness comes from a person's path through the world: the problems encountered, the questions pursued, and the corrections accepted. Distinctive does not mean correct. A highly personalized system can also preserve a highly personalized mistake.
“Optimal knowledge” therefore needs an objective. For the owner, a useful objective is better decisions and learning per unit of attention, with privacy and maintenance costs included. For a buyer, it is better task outcomes at acceptable cost and risk. Those objectives overlap, but neither implies storing everything.
Start with one domain. Record ten meaningful cases: context, decision, alternatives, evidence, result, and later corrections. Keep several unfamiliar cases aside. Compare the individual agent with the same general model without the individual records. Score correctness, useful abstention, source accuracy, and whether a human can verify the result. Ten cases can expose defects; they cannot establish broad reliability. Keep testing as the claimed scope grows.
Keep original records portable and search indexes disposable. Add semantic retrieval when ordinary search misses relevant material. Test model changes against the same cases. A fresh embedding or a fluent summary is not evidence that the system has learned something true.
External evidence · What exists in September 2026
individual storage: Solid describes user-controlled Pods and interoperable access by applications. This is a concrete storage and permission pattern. It does not establish a market price for someone's knowledge. Solid overview.
Agent exchange: A2A specification v0.3.0 defines Agent Cards, service discovery, task exchange, and authentication-related behavior. This is a versioned protocol example, not a claim that it is the latest version. The standard is not proof of service quality, payment enforcement, or protection against learning from outputs. A2A specification.
Embodied capability and limits: Google DeepMind's July 2026 On-Device 2 model card describes local vision-language-action inference and restricted tester access. It also identifies limitations in unfamiliar tasks and high-degree-of-freedom control, and distinguishes its evaluated manipulation scope from mobile and whole-body risks. The wheeled workshop robot in the story is a speculative integration, not a demonstrated capability of this model. Model card.
Industrial reality: BMW's March 2026 account describes a humanoid pilot at Leipzig and prior experience at Spartanburg. It supports the existence of bounded industrial deployment activity. It is a organization report, not independent proof of general autonomy or profitable small-workshop operation. BMW report.
Human differences: Anthropic's March 2026 study associates longer user tenure with more successful and complex Claude use. It discusses cohort effects and survivorship bias. This supports investigating learned human–AI working practices; it does not prove that individual agents create income or that experience causes all observed differences. Learning curves report.
Quality uncertainty: The Royal Swedish Academy's 2001 economics explainer discusses asymmetric information, including Akerlof's work on markets where buyers cannot distinguish quality. That established problem motivates the testing mechanism below; it is not direct evidence about a future agent market. Economics prize background.
These sources establish components and constraints. None demonstrates the complete person-owned knowledge rental economy imagined here.
Codex inference · What an organization recruits
An organization would recruit a capability with an accountable provider: “assess this class of failure, under these conditions, with this evidence and response time.” A personality replica is harder to evaluate and usually asks for more access than the task requires.
The buyer submits a case and receives a result, supporting evidence, uncertainty, a version identifier, and a usage receipt. The owner's service retrieves only permitted knowledge. The organization combines that result with its own records and acceptance tests. New learning returns to the owner only under agreed permissions; client secrets do not automatically become the provider's individual property.
Technical possession, privacy rights, copyright, contractual permission, and rights in work-related material are different questions. This scenario assumes permission has been established; it makes no jurisdiction-specific claim that a worker owns everything they have encountered.
An illustrative service agreement would specify scope, retained outputs, training and redistribution permissions, expiry, concurrency, availability, acceptance criteria, dispute review, and responsibility for failures. These are design requirements, not a legal template. Revocation stops future authorized access. It cannot reliably erase copies, derived knowledge, or already trained models.
Consumption pricing needs a meaningful unit. Token charges cover computation but reward verbosity if treated as value. A more useful proposal combines an availability fee with a capped charge per accepted assessment or work unit. Robot time, hardware, operator supervision, and energy remain separate costs. Retries carry task identifiers to avoid duplicate billing.
Payment solely for “prevented failures” is also weak: the counterfactual is usually unobservable. In the story, Mai earns a fee for an independently accepted exception report, not an invented percentage of a catastrophe that never happened. Audits sample whether stops were warranted; otherwise her agent could earn more by stopping everything.
The organization still earns its place by integrating services, verifying results, maintaining continuity, operating equipment, and taking responsibility for customer commitments. Some knowledge is too tightly coupled to internal systems to rent efficiently. Employment, internal agents, service contracts, and open shared knowledge can coexist.
Codex inference · Five incentive tests
1. Why would anyone buy an unknown person's agent?
If low-quality and high-quality agents make identical claims, a buyer discounts both. Good contributors can leave when the discounted price does not cover validation and maintenance. The remaining catalog gets worse. This is the proposed agent-market analogue of adverse selection.
Useful responses include paid trials on undisclosed cases, verified outcome histories with consent, independent checks, and a known provider who can be contacted when results fail. A badge or self-reported benchmark can be gamed. Evaluation must include representative failures and tasks the service should decline.
2. Why would the owner keep contributing after being copied?
Consider a simplified repeated relationship. These are invented payoff points, not currency or empirical estimates. Respecting the license gives the buyer 6 each period. Copying gives 10 now, a one-time gain of 4. If detected, the relationship ends and a fallback provides 1 per future period. Let δ represent how much the buyer values the next period.
With certain detection and no additional sanction, respecting the agreement beats copying when:
6 / (1 − δ) ≥ 10 + δ / (1 − δ), so δ ≥ 4/9 ≈ 0.444.
Future cooperation can outweigh a quick gain. This conclusion assumes a credible termination, persistent service value, and the stated fallback. If the copy yields 6 forever, the threatened loss disappears. If the owner's old cases rapidly become generic, the recurring value also shrinks.
For the reader's interactive variant, let p be the chance that this one deviation is detected, F an additional collectible sanction, and assume detection triggers the permanent loss above. With no later discovery after a missed deviation, the incentive condition is:
4 ≤ p × [F + δ × (6 − 1) / (1 − δ)].
At p = 0.5, δ = 0.5, and F = 4, expected cost is 4.5, above the gain of 4. With F = 0 it is 2.5, so copying pays. This tests one buyer decision under specified assumptions. It is not a full market equilibrium, and detection or collection may be expensive.
The owner also needs a reason to participate: receipts after platform and service costs must cover maintenance, exposure, and forgone alternatives. The buyer must prefer the net result to internal work or another supplier. A license-respecting buyer does not, by itself, make the service worthwhile.
A buyer gains 4 points by copying once. Change the assumptions to see whether its expected cost is higher.
Respecting the license pays under these assumptions.
Expected cost = p × [F + δ × 5 / (1 − δ)]. Assumes detection ends the relationship permanently, a collectible sanction, and a lower-value fallback. This is one incentive check, not a prediction or a complete market equilibrium. At equality the buyer is indifferent.
3. What stops the platform taking most of the income?
A provider's bargaining power depends on outside options. A portable archive helps, but does not move customer relationships, reputation, test results, or payment access by itself. A platform can allow data export while making commercial departure costly.
The cooperative improves an outside option by sharing testing and customer acquisition. It works only if its benefits exceed its fees and governance costs. It can also become an exclusionary gatekeeper. Members need transparent allocation rules, appeals, viable exit, and tests that admit newcomers rather than protecting incumbents.
4. Who gets paid when several agents solve one problem?
Contribution is not token volume. A two-word objection may matter more than a long analysis. Paying by claimed savings invites inflated claims; splitting equally can reward free riding; paying only the final agent erases supporting work.
A plausible early market uses negotiated shares by work package and measured performance on agreed cases. Audits and disputes remain necessary. There is no automatic, universally fair way to price joint knowledge. When coordination and verification costs exceed the value gained, the organization should use fewer providers or bring the capability inside.
5. Why would society replenish the knowledge it rents?
Established experts have cases, equipment access, and reputations. Beginners have fewer of all three. A pure usage market can reward past opportunity and underfund the experience that produces the next generation of knowledge.
Paid apprenticeships, shared test facilities, public reference knowledge, and cooperative learning funds are possible responses. They require durable funding and governance. Society should also protect the ability to keep a individual system private. Universal access to a useful memory does not require universal participation in a marketplace.
Codex inference · Robotics changes the bargain
Software advice can be copied cheaply; bodies remain constrained by location, wear, calibration, energy, and supervision. A field case can therefore stay valuable even when a generic explanation becomes cheap. That is an opportunity, not a guarantee: robotics providers can also accumulate proprietary experience faster than individuals.
A plausible design separates advice → approved task plan → robot-specific execution → independent safety controls. The individual agent supplies context or a tested skill package. It does not gain unrestricted motor authority because a customer paid for access.
Licenses for physical work would need to bind use to a validated robot configuration, environment, and task. An offline package trades availability against revocation and copying risk. New embodiments require new evidence. Cheap language-model inference does not make repair dexterity, field support, or liability cheap.
For 2028–2030, constrained inspection and repetitive manipulation are more plausible entry points than a general robot adopting any person's expertise on demand. The evidence above supports this direction; deployment cost and transfer reliability remain open.
Codex forecast · 2028–2030
These are qualitative forecasts as of 14 September 2026, not measured probabilities. Confidence concerns the specific milestone, not universal adoption. The rows describe developments that could coexist.
| Horizon | Forecast and confidence | Conditions | What would weaken it |
|---|---|---|---|
| By end of 2028 | High: useful individual memory systems coexist with narrow paid expert-agent services in some domains | Retrieval, controlled access, recognizable task value | individual context adds little against improving general models, or setup costs dominate |
| By end of 2029 | Medium: some organizations procure external agents with trials, service terms, and human escalation | Repeat demand and affordable verification | Buyers prefer bundled platform agents or integration takes more effort than it saves |
| By end of 2030 | Medium: bounded robot workflows use externally supplied domain knowledge or validated skill packages | Task-specific validation, local controls, economic hardware deployment | Transfer failures, support costs, or accountability disputes overwhelm savings |
| By end of 2030 | Low: a broad cross-platform market gives most people meaningful agent rental income | Portable reputation, clear permissions, buyer diversity, low overhead, accessible learning | Platform concentration and commoditization absorb the gains |
The leading scenario is a mixed economy: many people use individual systems; fewer sell services; organizations rent selected capabilities while retaining internal teams and agents. A more hopeful branch has interoperable marketplaces and cooperatives. A less hopeful branch has portable notebooks but captive reputation and pricing controlled by a few platforms.
Track repeat purchases after trials, net provider earnings after maintenance, the cost of moving reputation between platforms, evidence of successful provider switching, novice access to real work, and robot performance outside its initial site. Those signals say more than registrations, token volume, or impressive demonstrations.
Open questions
Can a individual service outperform a strong general model on unfamiliar cases, after verification costs? Can buyers obtain continuity without demanding unlimited reuse? Can contributors carry credible reputation elsewhere without exposing customer information? Can a beginner acquire experience before already having a marketable agent?
The first practical experiment is small: one person, one domain, portable case records, a bounded agent, and an honest comparison with the generic alternative. If it helps only its owner, it can still be worth building. If it helps others, test repeat demand before imagining a passive income stream.
Mai's defensible asset is not everything she remembers. It is the continuing ability to notice, test, correct, and make those judgments available on terms she can accept.
Sources
- Solid: About — storage and access pattern; live page checked 2026-09-14.
- A2A specification v0.3.0 — versioned communication standard; checked 2026-09-14.
- Gemini Robotics On-Device 2 model card — published 2026-07-30; developer-reported capabilities and limits.
- BMW: humanoid robotics at Leipzig — published 2026-03-09; organization-reported deployment activity.
- Anthropic: Learning curves — March 2026; observational platform research.
- Royal Swedish Academy: 2001 economics prize background — background on asymmetric information; historical theory, not a forecast.