Agency and the constructed self
How we revise the rules we live by: bounded agency, reflective freedom and the constructed self
Jeremy Rodgers | Shadow Theory
Agency begins within conditions
A person who declines a cup of coffee because it will disturb their sleep has made a decision with consequences. Remembering, anticipating and weighing reasons contribute to what happens next. Those capacities have a history. The person did not choose the beginning of that history, the body through which they act, or all the circumstances in which the decision arises.
This is the starting point for an account of agency that takes both action and dependence seriously. People can learn, deliberate, resist impulses and change their conduct. None of these achievements makes them the intentional creator of everything that enables them to choose.
Shadow Theory places these activities within the readout: the differentiated world of bodies, relationships, instruments, memories and lived perspectives. Awareness is fundamental in the framework, but it is not an additional person standing behind the organism and directing it. Choosing belongs to organized activity. The question is how that activity works, what it can control, and what role its representation of itself plays. [1]
What a choice can change
Agency is always exercised somewhere, through particular capacities, over particular outcomes. A person may decide what to eat without having chosen the food available, the conditions under which it was produced, or the appetite they bring to the table. Limited control over one feature of a situation does not confer control over the whole situation.
The distinction also matters under coercion. Selecting between two imposed outcomes need not amount to consenting to the conditions that made those the only options. A description of agency should identify what the selection actually changes, rather than treating the existence of any choice as evidence of unrestricted freedom.
Information and control are different capacities. An agent does not need a complete description of the world to act successfully. It needs enough information to handle the distinctions relevant to its task. If two situations look identical to the agent and require incompatible actions, a rule based only on that appearance cannot guarantee success in both. If the same action works in both situations, the missing distinction may be irrelevant.
This gives bounded agency a precise meaning. Its limits depend on the task, the available information, the actions the system can execute and the resources it can use. More knowledge may improve control, but knowledge alone does not supply an actuator, time to deliberate or the capacity to carry out a decision.
Revising the rule of choice
An agent can change the rule it uses to decide. Bounded Agency and Reflective Freedom gives this claim a finite, testable construction. The controller retains commitments, assesses a small set of represented rules against them, and installs the best-supported rule for later use. The assessment can also leave the existing rule in place when it already meets the criterion. Reflection does not require constant change. [2]
In the paper’s example, the controller starts with a rule requiring both considerations to support an action. Assessment of its retained commitments changes the rule to accept either consideration. The current action register stays untouched during revision. A fresh case then receives a different answer because the installed rule has changed. This establishes a changed disposition, extending beyond a command written for the present occasion.
The construction separates three questions. Did the amendment agree with the declared commitments? Did reading those commitments actually influence assessment? Did installing the assessed rule change what happened on later cases? Interventions on the read route and the installation route test those causal roles. An outward actuator can fail even when an internal rule was successfully revised.
Causal influence by itself is insufficient. A controller that systematically installs the opposite of the endorsed rule can show the same strong causal contrast as a faithful evaluator. The endorsement relation must be checked separately. Equally, a lucky correct answer does not establish that the controller used its retained commitments to reach it.
This is a demonstrated capacity for proximal evaluative self-governance: the agent’s installed evaluation governs a later disposition. Its candidate rules are finite, its commitments are specified, and its method of assessment remains fixed. Revising a decision rule does not mean originating the standards by which that rule is assessed. The construction establishes the local capacity without demanding that an agent first choose everything that makes evaluation possible.
Spending the next question
Reflection has a cost. Evidence must be acquired, memories inspected and amendments installed before the opportunity to act passes. An acceptable rule can exist without being discoverable or installable through the agent’s available operations and remaining resources. The paper makes those restrictions part of the result. It also gives conditions under which further revision preserves the ability to assess and revise again. [2]
One exact model asks the controller to divide a common query budget between two tasks: calibrating an uncertain sensor arrangement from unlabeled observations and reading three retained endorsements to determine their majority. Success requires both a correct prediction about a fresh target and a correctly installed endorsement majority. Each calibration observation or endorsement read costs one query. The final target arrives only after inquiry, and unread endorsements are unavailable through any free channel.
The best use of the next query depends on what earlier queries revealed. If the first two endorsements agree, their majority is already known. The controller can spend the saved third read on calibration. If they disagree, it needs the third endorsement to settle the majority. Adapting the allocation makes a precise difference.
The first strict advantage occurs at a total budget of L + 2 queries, where L is the first calibration count that can improve target prediction in the specified sensor family. Before that budget, adaptive inquiry and the best fixed-count allocation have the same optimal success probability. At the threshold, the paper proves the best possible adaptive value over all admitted query orders, stopping rules and choices of unread endorsement. Random mixtures of fixed-count allocations are included in the comparison.
In the paper’s threshold-two example, L is two and the first advantage arrives at four queries. Agreement leaves two queries for calibration; disagreement uses a third endorsement read and leaves one calibration. The optimal gain over every fixed-count allocation is about 0.33361 percentage points. Another worked instance first gains at six queries. These are exact results for a declared finite family with specified sensor laws and independent, fair endorsements, not a universal four-query law for deliberation.
The allocation model uses a fixed majority charter. It demonstrates how inquiry can govern the use of limited resources; the earlier editable-rule construction demonstrates revision of the charter itself. They establish different capacities. Their achievements belong together in an account of agency, while their hardware and resource guarantees remain distinct.
Conditioning does not make deliberation idle
The accumulated organization through which a person interprets and acts can be called scaffolding. It includes learned concepts, memories, habits, skills and expectations. It works with bodily needs, present feelings and incoming evidence. Scaffolding is continually revised through experience, including the consequences of the person’s own actions.
A decision can therefore be conditioned and causally effective. Under determinism, the outcome occurs through the actual processes that produce it, including deliberation where deliberation matters. It does not occur regardless of those processes. Learning that a food is unsafe can alter the reasoning and prevent the person from eating it, even in a deterministic account.
Indeterminism introduces a different possibility: the complete prior conditions may permit several continuations. That openness alone does not establish freedom. An unpredictable interruption and a controlled act are different things. Equally, admitting uncertainty does not erase every form of control already present in a skilled or responsive process.
The position developed here is that local agency is real and conditioned. Its scope can grow through learning and shrink through restriction or impairment. Complexity can extend what an agent can do without making the agent the creator of its own foundations.
Will and thought
Differences do not all call for action. A discrepancy becomes relevant through its relation to needs, consequences, commitments and available possibilities. What matters to a person already depends on how their life is organized. An unfamiliar sound may be ignored, investigated or treated as a warning.
Within the developmental account, will is directed commitment toward reconciling a relevant discrepancy. Thought is the representational work attempting that reconciliation. Someone can remain committed to understanding a problem while every explanation they try fails. The direction persists even when the reasoning is unsuccessful. Thought can also proceed without deliberate willing, and directed regulation can precede explicit reasoning.
This distinction gives will a role without making it an uncaused force. Its direction can be shaped by acquired values and present conditions. Bodily needs and feelings can change the goal itself, the interpretation of a situation and the options considered. They are not merely interference added to an otherwise complete rational chooser. [3]
Experience changes the conditions of choice
Scaffolding does more than supply reasons for selecting from a fixed menu. Learning can change which possibilities a person recognizes and which actions they can perform. Understanding a disagreement differently can make a question available where previously there seemed to be only a choice between submission and resistance. Acquiring a skill can turn a previously impractical intention into an executable action.
The process feeds back into itself. A relevant difference directs inquiry. Thought constructs a possible resolution. If some consequence of that work is retained, it can change the organization through which the next situation is encountered. The person may notice a different feature, ask a better question, revise a priority or abandon an earlier method. Experience can also entrench a mistake. Development does not guarantee improvement.
Agency can therefore work in three ways: revising an interpretation, changing access to evidence, and acting on the world itself. A question changes what can be learned. A tool changes what can be measured or done. An institution can preserve earlier assumptions in rules that shape later experience. The agent participates in changing the conditions from which future choices arise. [3]
This also explains why equal performance today need not imply equal capacity to learn tomorrow. Two systems can produce the same current answer while retaining different distinctions. The same next input can then change one and leave the other unchanged. The relevant question is what their retained organization enables across further encounters, rather than what a single score reveals.
Conscious scaffolding
The consciousness proposal concerns how these acquired relations participate in experience. Recognizing a sound as part of a melody, a gesture as a warning or a familiar situation as a recurring mistake can change the organization of what is experienced. Within Shadow, this is a proposed change in conscious organization, while fundamental awareness remains unchanged.
The formal connection is conditional. Once a physical system meets the framework’s requirements for consciousness, its A2 principle links experiential organization to how the system can respond through its own physically available operations and records. If a retained change makes the possible recorded outcomes or their probabilities different, the framework assigns a different experiential state. That does not necessarily mean a new subject or a new overall organization of consciousness. A better task score alone does not establish the required internal change. Applying the result to a person or an artificial system requires evidence that the framework’s conditions are met. [3]
The agency paper supplies an additional construction: specified evaluators can be realized while preserving their computations, diagnostic interventions and restrictions on executive access. A physical record may exist without becoming a free observation for the decision procedure. Preserving that distinction is essential to preserving the inquiry bound. Matching ordinary answers alone would leave the causal and information-access claims unchecked. [2]
The realization also supplies the native return organization required for a conditional connection to SPC-2. Recurrence alone does not establish evaluative agency: the paper constructs recurrent organization without the relevant evaluative competence. Its connection to experience depends on the declared SPC-2 correspondence laws and the specified physical boundary. The result establishes compatibility within that framework, rather than empirical proof of consciousness in a device or person.
The account we give of choosing
The processes producing an action, our access to those processes and the explanation we subsequently give need not coincide.
In the original choice-blindness experiment, people selected a face from a pair. On some trials, the experimenter covertly substituted the rejected face. Participants often failed to notice and supplied reasons for the outcome presented to them. Some reports referred to features of the substituted face; others retained information about the original choice. The experiment established that explanations of choice can be mistaken. It did not establish that every explanation is invented, and the authors expressly left the mechanism of the effect unresolved. [4]
Split-brain research provides a different kind of evidence. Under particular testing conditions, information guiding a manual response can be unavailable to the system producing the verbal account. The relation between these functional divisions and the number of conscious subjects remains contested. A collective review in 2020 concluded that the evidence then available did not settle whether consciousness itself was divided. [5, 6]
Dream research makes another distinction visible. In a 2021 experiment, some participants answered previously unknown questions during verified REM sleep, using eye movements or facial contractions to communicate. The responses provided evidence of ongoing cognitive activity beyond a report reconstructed after waking. A brain-generated scene can coexist with responsive reasoning; being mistaken about the setting does not make every mental operation within it fictitious. This supports studying agency during dreaming without settling how a particular choice originates. [7]
These findings require care when interpreting the statement “I chose this.” The report can identify a real action while misidentifying some of its causes. It cannot serve as an infallible certificate of complete self-knowledge. Nor does an error in reporting show that the action had no organized cause.
Personal information and personal authority
A self-model contains information about the person: bodily capacities, memories, responsibilities, skills and expectations. Much of it is indispensable. Knowing that a movement will cause injury, or remembering a promise, can provide a sound reason for action.
A further question concerns the authority assigned to that information. A description such as “I am a capable person” may remain open to examination, or it may become a requirement that every interpretation must protect. The same content can play different regulatory roles.
Revising the description does not necessarily alter that role. A person can replace one story about themselves with another while continuing to require that some personal story remain the final authority. The change may be substantial at the level of belief while leaving the organizing relation intact.
The RCO proposal concerns this distinction. Its question is whether personal representation remains useful content or becomes the compulsory centre through which interpretation and action must be licensed.
RCO and non RCO
RCO stands for recursively closed observer. These are terms of the Shadow framework, introduced here to describe a proposed organization of self-reference. They are not clinical diagnoses or established scientific classifications.
In a non-RCO, self-reference terminates on the constructed self-model. The personal model functions as the privileged anchor around which interpretation and regulation are organized. Such an observer has awareness, consciousness and agency. The term does not imply an absence of reflection, intelligence, kindness or practical competence.
In an RCO, self-reference is understood to close on awareness rather than on the constructed self-model. Personal content remains present and usable, but it no longer occupies that privileged position. Within Shadow, this is a reorganization of how the person is related to their own representation. It is not an increase in fundamental awareness, which the framework attributes to both.
“Closes on awareness” does not mean that an internal process discovers a special object called awareness and installs it as a new controller. Awareness is not an additional register, witness or instruction source. The phrase expresses the framework’s interpretation of self-reference without a privileged constructed identity as its final anchor.
This distinction preserves ordinary human life. Memory, planning and commitments can remain. Disagreement can remain. Refusing a mistaken demand can be appropriate. Relinquishing compulsory self-authority does not require accepting every suggestion, abandoning judgment or becoming passive.
The interpretation is stronger than the operational tests presently available. A test can investigate a particular authority relation without establishing the entire RCO condition. Neither a declaration about one’s awareness nor competent behaviour alone classifies an observer. The terminology names a hypothesis to be examined, not a status to confer on people.
The new bounded-agency constructions sharpen part of this discussion. They show that an assessed rule can be revised and that its installation can govern fresh decisions. They do not establish that self-reference has ceased to terminate on a privileged personal model. A non-RCO could possess the demonstrated evaluative capacities. The full RCO hypothesis concerns a further organization of self-reference and requires its own evidence. [2]
Why correct performance is not enough
A simple decision model exposes the difference between content and authority. Give a system a stored personal recommendation and a channel carrying a new proposal. An authority gate can either enforce the stored recommendation or allow the proposal to determine the action. The recommendation remains available in both cases.
Now let the gate enforce the recommendation only when it agrees with the correct task response. When it would conflict with the evidence, the gate permits the evidence-based proposal. The system can perform perfectly while retaining compulsory authority in part of its operation.
A suitable intervention can reveal what the ordinary task conceals. Retain the recommendation and present a contrary proposal after the gate has been committed. Compare the response with the gate present and with its authority withdrawn. Under the model’s specified conditions, the contrast exposes whether the recommendation has a mandatory licensing role.
This is a finite organizational result. It does not imply that every override is harmful: a contrary proposal can be wrong, and a safety constraint can be warranted. The distinction concerns what controls the response. Whether that control is justified requires another assessment.
Persistence matters too. A system may relinquish an authority role temporarily and later reinstall it. A successful test on one occasion cannot certify every later context. Applying these ideas to a person would require independently grounded measures of the content, its authority and its return through learning. The model supplies a question and a controlled comparison, not a neurological diagnosis.
What the tests can establish
The developmental research supplies a complementary test of retained organization. Identify what an interaction changes, remove access to the original interaction, and ask whether the effect persists. Where a lawful intervention is available, remove the proposed retained mechanism and then restore it. Loss followed by a specific rescue can support the claim that this mechanism contributes to later behaviour. It does not establish that the original experience was the only way to acquire it. [3]
Testing also has limits. In an exactly specified finite model, a suitably complete diagnostic set can determine all of its finite continuation laws from a preparation. Applying this result requires the model, the diagnostic coverage and the relevant error bounds. Agreement on a few ordinary tasks supplies no such guarantee. A rare discrepancy can remain almost invisible in short trials and become decisive over a longer run. These results help distinguish a demonstrated developmental change from confidence based on a few convincing exchanges. [3]
Freedom from the constructed centre
Psychological freedom concerns how the constructed centre participates in perception and action. A person can select food, use knowledge and respond to circumstances without requiring every response to defend an image of themselves. The philosophical possibility is that practical judgment remains while the personal image loses its compulsory authority.
The RCO proposal gives this direction an organizational interpretation. A personal description can inform action without having to authorize every possible correction. A thought about oneself can remain available without being treated as an exhaustive account of what one is.
Clinical research on decentering offers a related, narrower comparison. Decentering concerns how people relate to thoughts and feelings as mental events. It should not be treated as an interchangeable name for RCO. Likewise, a 2022 study found better choice-manipulation detection among experienced meditators than among non-meditators, but its correlational design did not establish that meditation caused the difference or identify the responsible mechanism. [8, 9]
These comparisons help identify questions for investigation. They do not prove that reduced self-authority always improves performance, makes someone morally superior or produces a permanent transformation. Clear perception can still encounter inadequate information and genuinely conflicting needs.
The remaining question about free will
Complete intentional self-authorship is unavailable to a person with a well-founded developmental history. An intentional act already requires capacities through which it can occur. Earlier acts may have helped form those capacities, but tracing that process backward eventually reaches an enabling basis the person did not intentionally establish. Extensive self-change remains possible; complete authorship of its entire basis does not follow. This is the limited conclusion to retain from the self-creation regress. [10]
A further question remains. Could an acquired person exercise a capacity that genuinely settles a particular choice left open by the complete past and laws? That is the stronger libertarian claim under discussion. It does not require the person to have created the capacity, and the absence of self-creation therefore does not refute it by itself.
Neither randomness nor a feeling of authorship establishes such a power. But describing a process with probabilities does not, by itself, prove the power impossible either. A conclusive answer requires an account of what it means to control the settlement of an open alternative and why the relevant process does or does not satisfy it.
This question remains unresolved here. The agency paper proves that specified capacities for evaluation and represented-rule revision can coexist with deterministic implementation. It does not provide alternative outcomes from one identical complete deterministic state, settle libertarian free will, or decide moral responsibility from present architecture. It also makes no empirical claim that human decisions are deterministic. [2]
The history of the agent remains a further question. A present evaluator can follow its retained commitments even when those commitments arose under manipulation. Preserved internal assessment does not reveal whether someone controlled the evidence reaching it. Historical investigation therefore needs records of acquisition and source provenance, alongside tests of current decision making. The paper specifies how to examine those routes over a recorded interval without claiming to certify the whole of a person’s development.
These limits leave substantial work already achieved: choices can depend on unchosen conditions, deliberation can be causally effective, rules can be assessed and revised, and the authority of a constructed self-model can be distinguished from its useful content. Psychological freedom from compulsory self-authority does not have to wait for a verdict on every possible causal interpretation of choosing.
Acting within the world
Shadow’s account places the person who deliberates within the world they affect. Learning, conversation, effort and insight can change the organization through which action occurs. The person can help shape their future capacities without having intentionally created their original capacity to learn.
The practical question is what can be understood and changed: which differences become relevant, which information reaches a decision, which reasons can be considered, which commitments can be revised, and whether a personal image must be protected before anything else can be heard.
Agency gives those questions consequences. The RCO hypothesis asks whether self-reference can be reorganized while useful personal capacities remain. Its philosophical significance lies in that possibility. Its scientific value depends on making the proposed change precise enough to distinguish it from a new story the self tells about itself.
Continue to Agency and Free Will for the complete mathematical treatment and its wider philosophical development, or read the published paper.
Sources and further reading
1. Shadow Theory and Consciousness, Chapter 9: Resources, continued operation, and agency. Read source
2. Jeremy Rodgers (2026). Bounded Agency and Reflective Freedom: Evaluative Revision, Information Limits, and Faithful Realization. Read source
3. Jeremy Rodgers (2026). Relational Development and Conscious Scaffolding: Conceptual Foundations, Strengthened Results, and Research Commitments. Read source
4. Johansson, Hall, Sikström and Olsson (2005). Failure to Detect Mismatches Between Intention and Outcome in a Simple Decision Task. Science 310, 116–119. Read source
5. Volz and Gazzaniga (2017). Interaction in isolation: 50 years of insights from split-brain research. Brain 140, 2051–2060. Read source
6. de Haan and colleagues (2020). Split-Brain: What We Know Now and Why This is Important for Understanding Consciousness. Neuropsychology Review 30, 224–233. Read source
7. Konkoly and colleagues (2021). Real-time dialogue between experimenters and dreamers during REM sleep. Current Biology 31, 1417–1427.e6. Read source
8. Moore and colleagues (2022). Association between decentering and reductions in relapse/recurrence in mindfulness-based cognitive therapy for depression in adults. Journal of Consulting and Clinical Psychology 90, 137–147. Read source
9. Lachaud, Jacquet and Baratgin (2022). Reducing Choice-Blindness? An Experimental Study Comparing Experienced Meditators to Non-Meditators. European Journal of Investigation in Health, Psychology and Education 12, 1607–1620. Read source
10. Galen Strawson (1994). The Impossibility of Moral Responsibility. Philosophical Studies 75, 5–24. The article above uses the self-creation regress without treating it as a proof against every form of responsibility. Read source