How AI agents reach the real world
Five integration layers an AI agent has to cross to get out of the purely digital world — and act in the real one.
Today’s agents are wired to tools. But tool access is not agency. An agent doesn’t only need models of how the world behaves. It needs models of how the world is organized.
The stack
Each layer is an integration problem, and each answers one question an agent must settle before it acts. Miss a layer, and the agent is guessing in a domain where guessing has consequences.
What can I access?
The layer of information systems and computational tools. Data, interfaces, memory, messages — everything an agent can read or invoke.
This is where essentially all of today’s AI agents live. It is the floor, not the building.
What can happen?
The constraints and causal dynamics of matter. An agent operating real assets must know what those assets can actually do — and what they will do next.
For an energy agent: batteries, PV systems, wind turbines, grid infrastructure, data centres.
Physics defines possibility.
Draws on World Models · Embodied AI · Physical AI
What may or must happen?
The rules that govern action. A technically possible action may still be prohibited, conditional, or valid only in one jurisdiction on one day.
Feasibility and legality are different questions. Agents today answer only the first.
Law defines permission.
Draws on Normative Multi-Agent Systems · Electronic Institutions · LegalRuleML · Rules as Code
What is it worth?
The layer where value is evaluated and exchanged. Economic autonomy is more than processing payments: it means holding positions, making commitments, and optimising under constraint.
An agent that can pay is not yet an agent that can trade.
Markets define value.
Draws on Agent-Mediated Electronic Commerce · Automated Negotiation · Mechanism Design · Commitment Theory
Who may act on whose behalf?
The layer that determines whether an action carries legitimate authority. An agent may be able to execute a transaction, and that transaction may be lawful — without the agent having been authorised to execute it for this principal.
Capability is not a mandate. This is the layer the industry has barely started building.
Institutions define authority.
Draws on Electronic Institutions · Institutional Reality · Authenticated Delegation · Agent Governance
Cross-cutting
Every layer can be wrong, stale, spoofed, or unverifiable. So trust cannot sit on top of the stack — it has to be established at each layer, for every action, and it has to leave a record.
Reality+Authority+Evidence
The stack, applied
One agent, one asset, one day.
“Maximise today’s economic return from this asset.”
A digital agent can already read the market, forecast prices, and produce a trading strategy. That strategy is not yet an action it is allowed to take.
What information is available?
Rich. This part is solved.
Can the action actually be executed?
Half the strategies the model produced are physically impossible.
Is the action permissible?
Some of what remains is legal. Not all of it.
Which permissible action is worth most?
Only now is optimisation a well-posed problem.
Is this agent authorised to do it?
Nothing in the previous four layers answers this.
Only the intersection of these five realities defines the set of actions this agent can legitimately execute. Every layer it cannot represent is a layer where it is not deciding — it is assuming.
A formal view
Four of the layers decide which actions exist for this agent at all. The fifth decides which of those is best — so it enters the problem as the objective, not as another constraint.
The agent maximises economic value…
a* = argmaxa Ueconomic(a)…over what the other four layers leave available.
A_digital digitally accessible A_physical physically possible A_normative normatively permitted A_institutional institutionally authorised An autonomous agent searches for the economically best action within the intersection of what is digitally accessible, physically possible, normatively permissible, and institutionally authorised.
Why this is not a new field
None of these five realities is an unsolved research question. Each has decades of work behind it — in separate communities, with separate vocabularies, largely unaware of one another.
What LLM-based agents change is that a single actor now has to operate across all of them at once. The opportunity is not to invent the layers. It is to connect them.
Elinor Ostrom’s Institutional Analysis and Development framework decomposed action situations decades before anyone was building software agents. Its elements map onto the stack almost one for one — which suggests the decomposition is not arbitrary.
| IAD concept | Agent Reality Stack |
|---|---|
| Biophysical conditions | Physical World |
| Rules-in-use | Normative World |
| Payoff rules | Economic World |
| Positions | Institutional World |
| Information rules | Digital World |
| Action situation | The agent’s decision |
This shifts the question from how to build more capable agents — toward the infrastructure that lets agents participate in the real economy at all.
Reading
The stack is a synthesis, not a discovery. These are the traditions it stands on.
Conant, R. C., & Ashby, W. R. (1970). International Journal of Systems Science, 1(2), 89–97.
The cybernetic result that an effective regulator must contain a model of what it regulates. The theoretical floor under the whole argument.
Ha, D., & Schmidhuber, J. (2018).
Agents learning compressed internal representations of their environment in order to predict consequences.
(2025).
Extends world models toward perception, planning, and physical action.
Vázquez-Salceda, J. (2004).
Connects values, norms, rules, and procedures into structures that can govern autonomous agents.
Dagstuhl Reports 8(4).
Formalises obligations, permissions, prohibitions, sanctions, and compliance.
OASIS.
Machine-readable legal norms, including jurisdiction, exceptions, precedence, and temporal validity.
OECD.
Rules as Code — government rules published in machine-consumable form.
OECD (2026), public consultation.
Machine-executable representation of legal conditions, hierarchies, deadlines, discretion, and consequences.
Journal of Institutional Economics.
Ostrom’s IAD framework — action situations shaped by biophysical conditions, rules, positions, information, and payoffs.
Singh, M. P. (1999).
Formalises one agent becoming obligated toward another — the bridge between contracts, norms, and action.
Guttman, R. H., Moukas, A. G., & Maes, P. (1998). Knowledge Engineering Review.
Software agents participating in commerce: discovery, merchant selection, negotiation, buying and selling. From 1998.
Cardoso, H. L., & Oliveira, E. (2007).
How autonomous agents participate in environments governed by roles, protocols, contracts, and sanctions.
South et al. (2025). Proceedings of ICML 2025.
Explicit infrastructure for authentication, authorisation, delegation, access control, and accountability.
World Economic Forum (2026).
Proposes an Agent Capability and Authorization Profile (ACAP) covering delegated authority, enforcement, and auditability.
Dazzi (2026).
Related but distinct: the Society of Agents asks how populations of agents are organised. The Agent Reality Stack asks which dimensions of reality a single agent must model to act legitimately.