codex
11 KiB
3. From Boundary to Autonomy
A wall separates, but it does not thereby act. A filter mediates, but it does not necessarily maintain itself. The passage from Markov boundary to agent therefore requires a concept of autonomy. Autonomy concerns not only how variables are partitioned at a time, but how a system contributes to the continued existence of the partition that distinguishes it from its environment.
This idea has deep roots in theoretical biology. Autopoietic theories characterize living systems as networks that produce the components and boundaries that recursively sustain the network. Enactive approaches treat cognition as sense-making by an autonomous organism rather than passive reconstruction of a pregiven world. Active inference supplies a complementary formal picture: systems occupying a limited repertoire of viable states act on their environments and update internal states in ways that preserve their organization. Volume 2 can be strengthened by treating its Markov blanket as the statistical manifestation of such self-maintaining dynamics.
The distinction between insulation and autonomy is crucial. A rock can be statistically distinguishable from its surroundings. A sealed container can maintain a strong physical boundary. A designed controller can preserve a target variable. None of these examples alone settles whether the system is autonomous. Autonomy requires that the processes inside the boundary participate in producing, regulating, or restoring the conditions under which the boundary and organization persist.
Counterfactual intervention provides a rigorous way to express this requirement. Let (b_t=(s_t,a_t)) denote blanket states, (c_t) internal states, and (\lambda_t) external states. Consider interventions on the environment. If internal dynamics make a distinctive contribution to restoring or preserving the boundary trajectory, then the system exhibits a form of autonomous control. One possible measure is
[ \mathcal{A}{T} = \mathbb{E}\left[ \sum{t=0}^{T} \log \frac{p(b_{t+1}\mid b_t,c_t,\mathrm{do}(\lambda_t))} {p(b_{t+1}\mid b_t,\mathrm{do}(\lambda_t))} \right]. ]
This quantity asks how much knowing the internal state improves prediction of future boundary states under environmental intervention. A high value indicates that internal organization contributes to boundary maintenance rather than merely covarying with it. The measure is only a proposal, and it would need normalization and empirical calibration, but it captures the correct explanatory direction.
Autonomy is not identical to resistance to change. Adaptive systems sometimes preserve themselves by changing. A neuron alters firing rates, a brain revises beliefs, and an organism moves to a new environment. The preserved quantity is not necessarily a fixed state but a set of viability constraints. These may include bounded temperature, energy availability, structural integrity, or a repertoire of sensorimotor capacities. The agent maintains a region of possible organization, not a frozen configuration.
This point reframes the free energy principle. The principle is often described as saying that self-organizing systems minimize variational free energy or expected surprise. Philosophically, the important claim is not that organisms seek low numerical values. It is that the persistence of a system implies a restricted distribution of states, and that action and perception can be modeled as maintaining occupancy within that restricted set. The normative language of "preferred states" arises from viability: some trajectories sustain the process, while others terminate it.
Normativity is indispensable to agency. A hurricane has organized dynamics and a recognizable boundary, but whether it has interests or goals is doubtful. A bacterium's movement toward nutrients is intelligible in relation to the continued viability of the organism. The difference is not simply complexity. It lies in the recursive relation between activity and the conditions of continued activity. An autonomous system's states can be better or worse for that system because they affect whether the system persists as the kind of process it is.
Volume 2's cortical focus complicates this account. A cortical column does not independently maintain its metabolism, vascular supply, or anatomical boundary. Its activity contributes to organism-level regulation, but its persistence depends on larger systems. If autonomy is required for agency, a cortical column may be only a component of an agent rather than an agent in its own right. This is not a fatal objection. It suggests that agency is hierarchical and that different levels exhibit different degrees or forms of autonomy.
Nested autonomy can be described without treating every subsystem as a full subject. Cells are autonomous in some respects while participating in organisms. Cortical circuits maintain functional regimes while depending on bodily regulation. Human beings maintain organismic viability while depending on social and ecological systems. The existence of dependence does not eliminate autonomy; complete independence would make interaction impossible. What matters is the pattern of reciprocal constraint and the degree to which a process contributes to maintaining its own organization.
The Intellecton framework is especially suited to this nested view because it already emphasizes recursive organization. An intellecton can be defined as a process whose boundary and internal integration recur across scales. But recursion must not become a license for indiscriminate attribution. Each candidate level should satisfy explicit criteria: approximate blanket separation, counterfactual boundary maintenance, persistence across perturbation, and irreducible internal organization.
This yields a stronger definition of minimal viable agency. A candidate system is minimally agentic when it has a statistically identifiable boundary and when its internal-active dynamics make a counterfactual contribution to preserving a viability region across a nontrivial range of external perturbations. This definition excludes accidental blankets and passive enclosures. It also allows degrees: systems can maintain narrower or broader viability regions, respond to fewer or more perturbations, and exercise more or less endogenous control.
The definition does not require consciousness. Autonomy and experience may be related, but their identity should not be assumed. Plants, immune systems, and simple artificial controllers may exhibit forms of autonomous regulation without satisfying stronger criteria for phenomenal unity. Keeping these concepts separate permits empirical progress. Researchers can test autonomy through interventions even when consciousness attribution remains contested.
Autonomy also introduces history. A system's capacity to maintain itself is shaped by prior adaptation, learning, and development. The current blanket is the product of a trajectory. Neural connectivity reflects evolution and plasticity; action policies reflect past interactions; bodily boundaries are continuously repaired. This historical dimension cannot be captured fully by an instantaneous conditional independence relation. It requires temporally extended models.
The transition from boundary to autonomy therefore changes the ontology of Volume 2. The basic unit is no longer a set of variables satisfying a factorization. It is a temporally extended organization whose activities sustain a recurrent factorization under changing conditions. The Markov blanket remains central, but it functions as a measurable signature of self-maintenance.
This reconstruction also clarifies the relation between internal and external states. Autonomy does not mean that internal states are sealed from the world. On the contrary, the blanket enables selective openness. Sensory states allow environmental influence; active states allow the system to alter environmental conditions. Agency consists in regulating this exchange. A perfectly isolated system would have no meaningful perception or action. A system with no mediation would have no distinct organization. The agent exists through controlled coupling.
The next question is whether such an autonomous process also possesses intrinsic causal unity. Volume 2 answers through integrated information. That move is promising, but it requires a separate analysis because recurrent organization, causal irreducibility, and phenomenal experience are not interchangeable concepts.
Degrees and Failures of Autonomy
One might object that autonomy is too demanding because accepted agents cannot directly repair every boundary. Humans depend on caregivers, infrastructure, and ecosystems; software agents depend on servers. The criterion should concern contribution rather than self-sufficiency. A process is autonomous to the extent that its internal-active dynamics participate in preserving or reconstructing the conditions of its continuation. Dependence is compatible with autonomy when the process regulates that dependence.
It is useful to separate constitutive autonomy from interactive autonomy. Constitutive autonomy concerns production and maintenance of organization. Interactive autonomy concerns regulation of exchanges with an environment. A cell exhibits both strongly. A cortical circuit may exhibit substantial interactive autonomy while relying on organism-level constitutive autonomy. These distinctions prevent a single score from concealing different achievements.
Autonomy is also vulnerable to narrow optimization. A controller can preserve a measured variable while destroying the larger organization that made the variable meaningful. Empirical measures should evaluate a viability region rather than one target and test novel perturbations rather than scripted recovery. Characteristic failures can reveal what a system was maintaining: seizure, addiction, and collapse expose hidden dependencies and competing regulatory loops. Failure analysis is therefore a central method for testing whether the proposed boundary is genuinely maintained.
This analysis changes the status of the active states in Volume 2. They cannot be treated simply as output variables. Their philosophical significance lies in closing a loop through which internal organization changes the conditions of its own future. An action that has no consequence for continued organization is behavior, but not evidence of autonomy in the relevant sense. Conversely, even simple action can be agentic when it reliably regulates viability under changing conditions. The decisive property is not sophistication but recursive causal contribution.
For neural systems, this means experiments must extend beyond isolated circuit dynamics. Investigators should test whether circuit outputs alter sensory inflow, bodily state, or larger network conditions in ways that stabilize the circuit's functional role. A cortical blanket demonstrated only under open-loop stimulation would be weaker than one sustained in closed-loop behavior. Autonomy is visible most clearly when the system is allowed to participate in making the environment to which it responds.