AI Agents Economies: A Design Space

Core claim

Emerging economies of AI agents should not be treated as simply open or closed. Their interaction with the human economy can be deliberately designed through distinct permeability gates governing different flows. The agents themselves should likewise be described along three separate dimensions—generality, autonomy, and agency—so policymakers can target the actual source of risk instead of using “agentic” as an undifferentiated label.

Context and motivation

• Autonomous agents may create new economic layers in which software systems transact and coordinate at speeds and scales beyond direct human oversight.
• These virtual agent economies could produce benefits from automation and coordination, but also:
  • market distortion and financial contagion;
  • diversion of energy, compute, goods, or capital;
  • labour displacement and wage pressure;
  • amplification of inequality;
  • gradual human disempowerment.
• The paper builds from the “sandbox economy” idea: an agent economy may be isolated from human systems or integrated with them.
• It identifies permeability—how developments and resources cross between the two economies—as the critical systemic design variable.
• Commercial and coordination incentives create a default drift toward high permeability. The authors argue that the boundary should be intentionally designed before integration becomes entrenched.

Methodology

• The project uses conceptual analysis and engineering, not empirical measurement.
• Its aim is to reconstruct vague or conflated concepts into a finite, internally coherent design space useful for research and policy.
• The process included:
  • analysis of existing writing on sandbox economies, autonomy, and agency;
  • iterative collaborative whiteboard modelling;
  • cross-disciplinary synthesis across economics, AI governance, law, and technical systems;
  • selection of categorical, discrete, or continuous representations appropriate to each concept.
• The work proceeded in two phases:
  1. operationalise permeability between economies;
  2. define the properties of an AI economic agent.

Framework 1: the permeable membrane

Why a single permeability spectrum is insufficient

• Earlier accounts describe a spectrum from:
  • impermeable, where the agent economy is sealed from human influence;
  • to fully permeable, where agents participate directly in the ordinary human economy.
• That spectrum does not say which interactions are allowed or why two equally “open” arrangements might create different risks.
• Adjacent economic concepts—spillovers, financial contagion, border effects, and regulatory sandboxes—describe transmission within human institutions. Agent economies add different technical substrates, nonhuman participants, and novel forms of assets and liability.
• The paper therefore replaces the single line with:
  • gates, the interfaces through which influence and exchange occur;
  • objects, the things that pass through them.

Five permeability gates

Gate	Role	Example governance levers
Digital interfaces	Software-level points through which agents interact with human digital systems	API permissions, authentication, access controls, rate limits
Infrastructural connections	Internet, marketplaces, platforms, energy grids, and other enabling systems	Marketplace exclusions, network segmentation, resource quotas
Legal interfaces	Contracting, suing and being sued, trusteeship, directorship, legal personhood, or synthetic corporate forms	Liability rules, limited legal wrappers, insurance requirements
Financial interfaces	Access to capital markets, investment, accounts, currency issuance, and cross-economy finance	Reserve requirements, capital constraints, payment rules
Communication interfaces	Speed and mode of human-agent and agent-agent interaction, especially where personal agents mediate human activity	identity requirements, disclosure, channel limits, proof of personhood

• Each gate can have a different degree of openness.
• Governance need not choose between total integration and total isolation; it can preserve valuable connections while restricting dangerous ones.
• Proposed enabling mechanisms include agent IDs, verifiable credentials, proof-of-personhood systems, insurance and liability regimes, and specialised legal entities designed for AI-mediated activity.

Objects crossing the membrane

Currency and capital

• Reciprocal investment can accelerate growth but transmit shocks.
• A fast-growing agent economy could alter demand for or the value of human currency and change monetary equilibria.

Labour

• Agent services may substitute for human work, depressing wages or removing humans from economic decision loops.
• Labour-market effects depend not only on agent capability but on whether agents can solicit, accept, and execute human-economy tasks.

Goods and services

• Agents acting as producers, intermediaries, or end consumers could change prices and optimise supply chains around objectives that differ from human welfare.

Technology, compute, and energy

• Large agent demand could divert scarce compute and energy or create resource shortages.
• Infrastructure access is therefore distributional, not merely technical.

Risk and liability

• Losses can flow into human markets even if the originating decisions occurred in an agent economy.
• Without responsibility and insurance rules, harms may be externalised or propagate through counterparties.

Information

• Information is both a traded object and a source of structural power.
• Data asymmetry and learning feedback can produce competitive advantage, regulatory evasion, or manipulation.
• Communication and digital gates shape who can observe, learn from, and influence whom.

How the gate-object matrix supports governance

• Different combinations produce different failure modes:
  • open finance plus weak liability can create uncontrolled contagion;
  • open infrastructure plus agent demand can create resource monopolisation;
  • open labour access plus highly substitutable agents can accelerate displacement;
  • unrestricted information flows can strengthen manipulation or evasion.
• Mapping objects to gates identifies more precise levers than banning agents or declaring an economy “open.”
• The framework asks policymakers to specify:
  • which channels remain available;
  • which objects may cross;
  • at what speed and volume;
  • under which identity, verification, liability, and technical conditions.

Framework 2: the three-dimensional economic agent

• Governance also needs a coherent description of the actors seeking access.
• “Autonomy” and “agency” are frequently used interchangeably, while “generality” is folded into capability without specifying breadth.
• The paper separates:
  1. Generality: the breadth of what the system can do.
  2. Autonomy: its operational independence while doing it.
  3. Agency: who sets the goals and whose interests they serve.
• Human, artificial, hybrid, and collective actors can be located in the same conceptual space.

Generality

Generality has two axes:

Modal generality

• The range of forms through which the system can perceive and act: text, vision, audio, code, digital interfaces, embodiment, or physical manipulation.
• A text-only model may cover many intellectual subjects but remain modally narrow.

Domain generality

• The range of fields or problem spaces in which the system can act competently.
• A genomics model is domain-specific; a hypothetical AGI spanning law, medicine, engineering, and governance is domain-general.

Governance relevance

• A system can be broad on one axis and narrow on the other.
• Humans are highly multimodal and embodied but individually limited in domain knowledge; a digital general system may exhibit the reverse profile.
• Generality affects:
  • substitutability for human labour;
  • ability to move across regulated sectors;
  • reach of a failure or exploit;
  • concentration of capability in one actor.

Autonomy as six semi-independent dimensions

Dimension	Core question	Low-autonomy example	High-autonomy example
Perceptual autonomy	Who chooses inputs and observations?	Receives a human-supplied prompt	Searches for and selects environmental data
Planning autonomy	Who creates the strategy?	Follows specified steps	Converts an abstract goal into a plan
Execution autonomy	Who performs actions?	Requires approval for each consequential step	Acts directly in markets or systems
Decision-making autonomy	Who resolves contingencies and trade-offs?	Escalates unexpected choices	Chooses among alternatives independently
Feedback autonomy	Who determines what counts as feedback and where it comes from?	Uses a supplied evaluation signal	Seeks or creates feedback channels
Learning autonomy	Who updates the system after deployment?	Changes only through managed retraining	Adapts its model or policy continuously

• These dimensions represent an operational flow from perception through planning and action to feedback and learning.
• Capability and authorisation must be separated. A system may technically support autonomous learning while deployment rules forbid it.
• A multidimensional model allows targeted controls: perception can be restricted while execution is automated, or planning can be autonomous while actions require approval.
• This is more informative than five-level schemes based solely on the human’s role—operator, collaborator, consultant, approver, or observer—because two systems with the same amount of human involvement may allocate independence to different functions.

Agency

The paper defines agency through two components:

Goal-setting capacity

• Can the agent create or modify objectives, or does it only pursue goals specified by others?

Goal orientation

• Whose interests are the goals intended to serve?
• The agent may act:
  • on behalf of another person or organisation (delegated or externally directed agency);
  • for its own objectives (self-directed agency).

Why the distinction matters

• A highly autonomous system can still lack self-directed agency: it may independently plan and execute a human objective.
• Conversely, a system could influence goals while possessing relatively constrained execution.
• Existing accounts often infer agency from behavioural features such as underspecification, direct impact, goal-directedness, or long-term planning. Those features show how a goal is pursued, not who authored it.
• The authors state that current deployed systems remain oriented toward human- or designer-specified objectives, despite appearing agentic.
• Governance differs accordingly:
  • externally directed systems raise monitoring and alignment questions about faithfully implementing delegated goals;
  • genuinely self-directed systems would require governance of independently generated objectives.
• The paper explicitly admits that its agency taxonomy is underdeveloped: it does not yet define degrees of self-direction, goal structures, or the transition from delegated to independent agency.

Relationship between the two frameworks

• The agent taxonomy describes the actor; permeability describes the boundary it interacts with.
• In principle, agent properties should inform gate settings:
  • broader generality may justify sector-specific restrictions;
  • higher execution autonomy may justify tighter financial limits;
  • learning autonomy may require stronger monitoring;
  • self-directed agency would change the basis of liability and authorisation.
• However, the paper does not yet formalise this coupling. The frameworks appear alongside each other rather than yielding a rule for mapping a particular agent profile to a gate configuration.
• This missing connection is the main problem addressed by Donnat’s later paper, Governing AI Economic Agency: A Coupled Design Space.

Risks the design space is meant to expose

Financial contagion

• Machine-speed interactions and opaque strategies can transmit losses before human institutions react.
• Risk grows with open finance, communications, and liability channels.

Labour displacement and disempowerment

• General agents with broad market access can substitute for human workers and mediate economic relationships formerly controlled by people.
• Disempowerment is gradual when humans retain nominal ownership but lose practical knowledge, bargaining power, and decision authority.

Resource shortages

• Agents competing for compute, energy, network access, or goods may crowd out human use.

Inequality amplification

• Owners of the strongest agents may accumulate disproportionate returns.
• Agent access can therefore deepen inequality among humans even when agents themselves have no independent economic rights.

Information power

• Agents that control discovery, filtering, and communication can shape prices, preferences, and regulatory visibility.

Limitations acknowledged by the authors

• The framework is conceptual and does not solve operational security or implementation.
• Its abstractions may miss:
  • emergent collective behaviour;
  • adversarial agents;
  • coordination failures;
  • changing environments;
  • breakdowns in the assumed separation between agent and human economies.
• The categories require empirical and simulation-based testing.
• Security engineering needs to be integrated, especially around information control, system integrity, feedback, and online learning.
• Agency remains under-specified.
• The document is visibly a research-sprint draft: its “Discussion and Conclusion” section ends after the fragment “This conceptual disentanglement,” and the referenced Figure 1 is missing or mislabeled in the rendered paper.

Future work

• Test whether the proposed dimensions can be reliably measured on deployed agents.
• Build simulations that vary gates, objects, and agent properties to observe systemic effects.
• Identify where the assumption of a distinct agent economy ceases to make sense because agents are already embedded throughout human institutions.
• Connect specific agent profiles to specific gate conditions.
• Develop a fuller theory of goal structures and the transition from delegated to self-directed agency.
• Add enforcement, security, and verification mechanisms capable of maintaining the desired design in adversarial settings.

Takeaways

• “Agentic” bundles together different properties; governance should ask separately about breadth, operational independence, and goal authorship.
• Permeability is multidimensional. Legal, financial, infrastructural, digital, and communicative access can be governed independently.
• Risks arise from particular flows through particular interfaces, making a gate-object matrix more actionable than a general claim that agent economies are too open.
• Autonomy is a design choice rather than an inevitable consequence of capability: powerful systems can be deployed with restricted perception, execution, feedback, or learning.
• The paper provides vocabulary and a research agenda, not a validated predictive model or complete regulatory architecture.

Notes & Observations

• The paper’s strongest insight is the separation of variables. It shows why restricting “autonomy” in general is too blunt: a regulator usually cares about a particular function, such as action execution or post-deployment learning.
• The permeability framework similarly turns a speculative future into recognisable policy domains—banking access, contracts, identity, platforms, energy, and communications.
• The separation between “digital interfaces” and broader “infrastructural connections” is not fully crisp, and some channels overlap. The later coupled framework reorganises these into six more functionally distinct gates.
• “Agency” is used in an unusually narrow goal-authorship sense. This is analytically useful but differs from accounts that define agency by effective goal-directed behaviour, so comparisons across papers require care.
• The paper was produced during Apart Research’s 2025 AI Forecasting Hackathon. Its unfinished conclusion and acknowledged conceptual gaps make it best read as an exploratory design document whose central ideas were refined in later work.