Organized Matter, New Causality

Constructing Agency from Control to Language

Dar Aystron

Independent Researcher

Abstract

What happens when we organize matter into an artificial agent?

Artificial intelligence is usually described through capabilities: perception, reasoning, memory, learning, planning, and language. This note proposes a different perspective. By organizing physical processes into a particular causal architecture, we create a new causal driver in the world: a system that selects actions, changes its environment, observes the consequences, and continues acting.

OODA Agency Theory (OAT) takes a constructive approach. Agency begins with a minimal control loop: Observe, Orient, Decide, Act, and becomes progressively richer as we introduce episodic and semantic memory, learning, needs, attachments, attention, and self-modeling. Each capability extends the causal organization through which the agent interacts with its environment.

We develop this argument using a thermostat that evolves into a home agent. We then compare software implementations that demonstrate the same external behavior but realize different internal causal structures. Finally, we introduce language as an additional capability embedded within an existing agent, offering a different perspective on Searle’s Chinese Room.

The hypothesis concerns the emergence of system-level causal powers through the deliberate organization of matter. It requires neither new physical laws nor an assumption of phenomenal consciousness.

1. The central hypothesis: organizing matter creates new causality

We ordinarily describe artificial intelligence in terms of what a system can do. A system recognizes objects, learns patterns, solves problems, generates language, or operates tools. These descriptions are useful, but they do not necessarily explain what we have created by assembling these capabilities.

The universe admits causal descriptions at multiple levels of organization. OAT takes a constructive approach: it specifies a particular higher-level causal organization: agency, and realizes it through physical and computational mechanisms. [4, 7]

We’re not simply creating additional software functionality. We’re organizing matter so that a new level of causal organization becomes physically realized and operationally effective.

Consider a thermostat. Its individual components participate in ordinary physical processes. When organized into a feedback controller, they collectively acquire a new causal capability: regulating the temperature of an environment. The room behaves differently because that particular organization of matter has been introduced. [1, 2]

No new fundamental force has appeared. But the assembled system has causal powers that its unassembled components did not possess as a system. [4]

We propose that artificial agency can be understood constructively in the same way. By implementing a particular causal organization, we create a system capable of producing selected interventions in its environment. Additional capabilities extend the scope and sophistication of that organization.

This is our starting point: agency is a physically realized form of organized causality.

2. Starting with a thermostat: minimal agency

A thermostat provides a simple example of closed control. We can describe its operation through OODA [8]:

Observe. Measure room temperature, including the effects of previous heating actions.

Orient. Compare the observed temperature with the desired range.

Decide. Select whether to activate or deactivate heating.

Act. Execute the selected action.

The loop continues. Changes produced by the action, together with other environmental changes, enter the next Observe phase.

The thermostat has extremely primitive agency. It controls one variable, has a tiny action repertoire, and operates within an almost entirely predefined causal structure. Yet it demonstrates the foundational mechanism: selected actions change the environment, and subsequent observations influence further action selection.

In OAT, operational closure can have different scopes and levels of sophistication. A thermostat maintains closure within a narrow regulatory domain. More sophisticated agents extend this closure by integrating richer perception, memory, learning, internal and external actions, and increasingly complex causal relationships with their environment. [7]

We therefore do not need to identify a mysterious moment when a machine suddenly acquires agency. We can investigate how increasingly rich forms of agency arise through successive extensions of a basic causal organization.

3. Extending the thermostat

Let us progressively extend the system without abandoning the original control loop.

3.1 Perception: recognizing people

We add cameras and person recognition. The system identifies Anna and Bob, learning that Anna prefers 23°C while Bob prefers 20°C.

Temperature regulation now depends on who is present. People become distinguishable entities with different causal significance for the controller’s actions.

3.2 Episodic memory: history enters causality

We introduce episodic memory. The system records particular events: Anna arriving home, Bob leaving, temperature adjustments, and their observed consequences.

It learns that Anna normally arrives around 6 p.m., but sometimes returns earlier on Fridays.

Previous experience can now influence future action. The system’s individual history becomes part of its causal organization. [7, 8]

3.3 Semantic memory and learning: acquiring knowledge

The system generalizes from experience. With additional weather information available, it learns how quickly the house warms under different outdoor conditions and which occupancy patterns recur.

These generalizations form semantic memory. Learning allows the system to revise its knowledge and improve future decisions.

Episodic memory preserves particular experiences; semantic memory preserves acquired knowledge about the world. Together, they allow an agent’s past to influence how it interprets and acts in new situations. [7, 8]

3.4 Internal sensing: observing its own condition

We introduce sensors for electricity consumption, operating costs, and equipment condition.

The system now perceives aspects of its own operation alongside external conditions. Combining energy consumption measurements with weather information, it can learn how outdoor conditions affect heating costs and anticipate that maintaining 23°C during peak electricity pricing may exceed its energy budget.

Internal state becomes part of the control problem, introducing a rudimentary form of interoception.

3.5 Needs: persistent regulatory demands

We introduce multiple needs concerning environmental conditions, resource use, continued operation, and the well-being of affected occupants.

In OAT, needs are implemented below cognition, but their current states are surfaced to cognition through interoception. The agent can reason about these states, evaluate possible actions, and select actions intended to satisfy its needs. However, cognition cannot simply rewrite need states. Changes require actions that affect the conditions governing those needs. [12]

Needs establish persistent regulatory conditions that influence orientation and action selection. They create a durable source of causal pressure within the agent’s organization.

3.6 Attachments: learned significance of entities

The system recognizes Anna and Bob, remembers its history with each of them, and evaluates how different actions may affect their respective needs.

Attachments connect entities to needs through learned relationships. Anna’s arrival can therefore have a significance that depends on the agent’s particular history with her, rather than merely selecting a predefined temperature preference. [10, 12]

These relationships can evolve through experience, changing how the agent evaluates subsequent situations and actions.

3.7 Attention: acting on perception

As sensory channels and internal models multiply, the agent must determine what deserves attention.

In OAT, attention is an internal action. The agent can select what to examine, retrieve particular memories, or investigate an unexpected event. [9]

This approach is related to earlier work on perception and action, particularly William Powers’ Perceptual Control Theory, which treats action as a means of controlling perception. [3]

The agent can therefore influence its own subsequent perception. This introduces a recursive causal capability: the agent’s actions help determine what information will enter future orientation.

3.8 Self-model: the agent enters its own world model

Finally, we introduce a self-model.

The agent learns about its own capabilities, limitations, and reliability. It may recognize that its occupancy predictions are inaccurate or that a sensor has become unreliable.

The agent can now include its own condition in its causal understanding and select actions intended to improve its performance.

Throughout these extensions, the organizing OODA process remains. What changes is the set of causal relationships participating in perception, orientation, decision-making, and action. [7, 8]

4. From thermostat to home agent

As capabilities expand, it becomes useful to separate the physical temperature controller from the agent’s intelligence.

The thermostat remains a sensor and actuator. Cameras provide visual information, energy meters report consumption, and other devices provide additional means of perception and action. The agent itself runs on a HomePod-like device.

The home agent maintains one agent-level OODA loop, integrating these capabilities within a persistent control process. [7, 8]

The agent’s identity is not necessarily tied to a particular physical device, but its hardware matters. Through experience, the agent learns causal relationships involving its sensors, actuators, and physical capabilities. Changes in hardware may invalidate some of this acquired knowledge, requiring the agent to adapt its models and learn how to operate with the new configuration. Its agency can persist through such changes, while its understanding of its own physical capabilities evolves.

This illustrates a central architectural distinction: the agent is the organized causal system, not any individual sensor, actuator, or computational component.

5. What makes the organization causal?

A conventional home automation system could implement many of the capabilities described above. It might recognize people, learn preferences, predict arrivals, and optimize heating.

The question is not whether we use conventional software. OAT can be implemented using conventional software. The question is how that software organizes causality.

5.1 Two implementations, two causal graphs

Consider two implementations that receive the same observation: Anna arrives home early on Friday.

Both set the heating to 23°C.

The first implementation uses a conventional automation sequence:

Anna arrives -> Person recognized -> Preference retrieved -> Heating rule evaluated -> Temperature adjusted.

The second implementation uses an OAT agent.

Anna’s arrival enters Observe. During Orient, the agent constructs its situation using episodic and semantic memories, needs, attachments, attention, and existing commitments. Candidate actions are generated and evaluated. Decide selects the primary action, and Act realizes an action bundle containing the primary action and subordinate actions. [8, 9]

The external result may be identical: the heating is set to 23°C.

But identical demonstrated behavior does not imply identical internal functionality and causal organization.

The conventional implementation may perform its heating operation without creating episodic history, evaluating attachments, or realizing persistent commitments. The OAT implementation includes these mechanisms within the causal process that produces its action.

This is not a distinction between two programming technologies. It is a distinction between the causal structures their implementations realize.

5.2 Action bundles and internal causality

In OAT, Act realizes an action bundle at the end of the OODA cycle. The bundle includes the selected primary action and subordinate actions, which may affect both the external environment and the agent’s persistent internal state. [9]

The bundle can realize changes involving several mechanisms:

Episodic memory preserves the episode as part of the agent’s history. Semantic memory may incorporate revised knowledge. Commitments are established or modified through realized actions. Attachments may change as relationships evolve. Attention may be redirected through an internal action. Goals and the self-model may also be modified through actions supported by the agent’s current situation.

The state of needs may also change as the agent’s conditions change, without allowing cognition to rewrite the underlying need definitions. [12]

These are not independent update handlers scattered throughout OODA. Changes to persistent agent state are organized through realized action bundles. [9]

Consequences of external actions, including effects not yet known at execution time, become available through subsequent observations. They may then influence the selection and realization of further actions. [8, 9]

This creates multiple causal relationships extending beyond the immediate external intervention.

5.3 Causality across episodes

Suppose the agent adjusts the heating for Anna and subsequently observes that the room warmed more slowly than predicted.

That observation may lead to a later action bundle that updates its semantic model. The revised model can influence future predictions and heating decisions.

Alternatively, a previous commitment to minimize electricity consumption may influence the agent’s evaluation of an action today.

These relationships operate across different timescales. Some affect the next OODA cycle; others remain relevant much later.

We can describe the resulting causal structure using time-indexed directed acyclic graphs (DAGs). The OODA process is recurrent, but its execution history can be represented as causally connected graphs across successive cycles.

The graph is not simply repeated with new inputs. Through learning and realized internal actions, the agent can modify the state, relationships, and models that influence its future actions. [7, 9]

Through realized action bundles, the agent changes both its environment and its own internal state. These changes, together with the consequences it subsequently observes, shape the causal organization that produces its future actions.

5.4 Behavior versus causal organization

The comparison reveals why external behavior alone may be insufficient to characterize an agent.

Two systems can demonstrate the same response to Anna’s arrival while implementing different internal functions. Conversely, conventional software that implements the same persistent OODA organization, including relevant needs, memories, commitments, and action bundles, would realize the corresponding form of agency.

The distinction is therefore not between traditional software and something beyond software.

It is between different physically implemented causal organizations.

Our interest is not merely in what a system demonstrably does, but in how its organization causes it to act and how that organization changes through experience.

6. The agent as a new causal driver

Before introducing the home agent, the environment contains people, physical devices, environmental processes, and various control mechanisms.

After introducing the agent, an additional organized causal entity participates in that environment.

It can initiate interventions based on its own ongoing control process. It may investigate an unexpected pattern, retrieve previous experiences, revise a prediction, request missing information, or act to satisfy an existing commitment without receiving a corresponding explicit command at that moment.

These interventions arise through the agent’s particular causal organization.

This is a system-level claim, not a proposal for new fundamental physics. The organization creates causal capabilities that can be described and investigated at the agent level while remaining physically realized through lower-level processes. [4]

The same physical event can be described at different levels. At one level, electrical signals cause a heating system to activate. At another, an agent turns on the heating because of its learned expectations and commitments concerning Anna.

Both descriptions can be causally informative. [4]

By organizing matter into an agent, we introduce a new causal driver into the world, one whose powers can evolve through its own experience.

7. Adding language: a new level of organized causality

We have constructed an increasingly sophisticated agent without requiring natural language.

Only now do we introduce speech recognition, language interpretation, and language generation. A large language model may provide substantial parts of these capabilities.

But language does not become the agent’s control architecture. It is embedded within the causal organization that already exists. [11]

Anna says, “I’ll be late tonight.”

The agent interprets her statement as an observation relevant to its current situation. The information can influence orientation, action evaluation, and the selection of an action bundle that updates its persistent knowledge or heating plan.

Now suppose the agent responds, “I’ll make sure the house is warm when you arrive.”

Within OAT, this can be a communicative action that establishes a commitment through action realization. The commitment becomes part of the agent’s persistent state and can influence subsequent OODA cycles. [9, 11]

Anna may rely on the promise, changing her own expectations and actions.

Language has therefore extended the agent’s causal reach. It can now ask questions, exchange information, negotiate, explain decisions, accept or refuse requests, and establish commitments that coordinate actions across agents. [6, 11]

The agent can also distinguish between hearing a statement and accepting it as true. Something Anna says may be preserved in episodic memory without automatically becoming an accepted semantic belief.

These capabilities are not merely transformations between input and output strings. They are linguistic actions and perceptions integrated with an agent’s needs, memories, relationships, and commitments.

Language becomes part of the agent’s organized causality.

8. Language Understanding and the Chinese Room

In 1980, John Searle introduced the Chinese Room thought experiment, challenging the idea that manipulating symbols according to formal rules is sufficient for understanding. [5]

Imagine a person who does not understand Chinese sitting inside a room. The person receives Chinese symbols, follows an elaborate rulebook, and produces responses in Chinese. To someone outside, the exchange may appear to demonstrate fluent understanding. Yet the person inside is simply following instructions, without understanding what the symbols mean.

Searle uses this example to argue that executing a computer program is not sufficient for intrinsic intentionality. He does not rule out artificial machines capable of understanding, but argues that they would need the relevant causal powers, not merely the ability to reproduce intelligent behavior through formal computation.

The Chinese Room has generated decades of philosophical debate, including the Systems Reply, the Robot Reply, and broader disagreements about computation, understanding, and consciousness. The fundamental question remains: what does it take for a physical system to genuinely understand something?

OAT approaches this question from a different direction.

We begin with physical control and progressively construct an agent with memory, learning, needs, attachments, attention, and a self-model. Each capability extends the causal organization through which the agent perceives, decides, acts, and changes through experience.

Consider two home automation systems. Both recognize Anna and set the temperature to 23°C. One uses a relatively simple program. The other is an OAT agent whose action emerges from an ongoing causal process involving its history, needs, learned relationships, and commitments. Its realized action bundle can also modify episodic memory, semantic knowledge, and other aspects of its persistent state.

The demonstrated external behavior is identical, but the internal functionality and causal organization are different.

From the OAT perspective, understanding can be investigated as a progressively developing causal capability. A thermostat detects and corrects temperature deviations, but it does not construct and revise acquired knowledge about its environment. Our home agent can predict how quickly a room will warm, recognize when that prediction is wrong, and revise its semantic model through subsequent actions.

The agent itself evaluates its predictions against subsequent observations and uses the discrepancies to revise its knowledge, without requiring an outside interpreter to assign those discrepancies significance.

This operational understanding becomes richer as additional capabilities enter the agent’s causal organization. Needs establish regulatory significance, attachments connect that significance to particular entities, episodic memory preserves experience, and attention directs further investigation.

Only then do we introduce language.

When Anna says “I’ll be late,” her statement becomes part of the agent’s perception and can influence its subsequent actions. When the agent promises to warm the house, the communicative action establishes a commitment through action realization. That commitment can shape future decisions and actions. This connects OAT to the theory of speech acts, in which language itself can perform actions. [6, 11]

Language is no longer an isolated process of transforming symbols. It is embedded within the causal organization of an agent with its own history, needs, knowledge, relationships, and commitments.

Searle’s challenge remains important: does this physically realized causal organization suffice for intrinsic intentionality? OAT does not claim to have settled that question, nor does it require us to assume phenomenal consciousness.

Instead, it offers a constructive approach to increasingly sophisticated operational understanding, investigating the causal mechanisms through which representations acquire significance for an agent and influence its actions.

Rather than asking how symbol manipulation alone becomes understanding, we investigate how language acquires causal significance within an agent whose needs, history, and relationships already shape what matters to its actions.

We do not derive agency from language. We embed language in agency.

9. Two paths toward the same architecture

The thermostat example follows a cybernetic path: begin with control and progressively add perception, memory, learning, needs, attention, and language. [1, 2, 7]

Contemporary AI development often follows another path. Begin with a language model, provide tools and an execution loop, and progressively introduce persistent state, environmental interaction, autonomous action selection, and commitments.

One path begins with control and adds intelligence. The other begins with sophisticated linguistic intelligence and adds persistent control.

Neither path guarantees the eventual architecture. An advanced controller may remain narrowly constrained, while an LLM with tools may remain a task-oriented automation system.

But both can potentially converge on a persistent agent-level causal organization.

This brings together several traditions in AI: cybernetics and feedback control, adaptive systems, knowledge representation, cognitive architectures, machine learning, and language models.

OAT does not propose replacing these traditions. It proposes a particular causal architecture through which their capabilities can become parts of one persistent agency. [7, 8]

10. Conclusion: constructing a new source of causality

OODA Agency Theory begins with a simple control loop and progressively extends its causal capabilities.

Memory makes history causally effective. Learning modifies acquired knowledge and future action selection. Needs establish persistent regulatory demands. Attachments give particular entities learned significance. Attention makes perception an object of action. Self-modeling allows the agent to include its own condition in its causal understanding.

Finally, language extends agency into communication, shared expectations, commitments, and coordinated action.

The larger hypothesis is that artificial agency can be investigated as a problem of constructing higher-level causal powers through the deliberate organization of matter.

This perspective does not require new physical laws or assume that increasingly sophisticated agency necessarily produces phenomenal consciousness.

It asks us to examine the causal entities we are capable of constructing.

The central question is not merely what an artificial system can compute, recognize, or say. It is what new causal powers we create when we organize matter into an agent and how those powers change as the agent acquires memory, needs, attachments, attention, and language.

We begin with a thermostat.

We end with an agent whose words, actions, relationships, and history can change what happens next in the world.

Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine.
Ashby, W. R. (1956). An Introduction to Cybernetics.
Powers, W. T. (1973). Behavior: The Control of Perception.
Ellis, G. F. R. (2016). How Can Physics Underlie the Mind? Top-Down Causation in the Human Context. Springer. Publisher.
Searle, J. R. (1980). “Minds, Brains, and Programs.” Behavioral and Brain Sciences, 3(3), 417–424.
Searle, J. R. (1969). Speech Acts: An Essay in the Philosophy of Language.
OAT Paper 01 — Operational Agency and the Architecture of Closure.
OAT Paper 02 — OODA as Agentic Control.
OAT Paper 04 — Action Realization in Agentic Systems.
OAT Paper 07 — Attachment-Based Ethics for OODA Agents.
OAT Paper 10 — Communicative Agency: Speech Acts Within OODA Control.
OAT Paper 11 — Needs and Motivation in OODA-Based Agents.