Haueis and Colaço argue that metabolic considerations should play a central role in cognitive modeling, distinguishing between evaluative uses of metabolism, which function as explanatory constraints on models, and generative uses, which help formulate new models of cognitive capacities and neural systems. In this commentary, we extend their argument to a theoretical commitment that is often implicit in cognitive modeling but rarely scrutinized through a metabolic lens: the assumption that cognitive capacities can be cleanly partitioned into relatively autonomous components (see Barrett & Kurzban, Reference Barrett and Kurzban2006, for a review). Strong modular views have been most clearly articulated in evolutionary psychology and classical modularity theory (Barkow, Cosmides, & Tooby, Reference Barkow, Cosmides and Tooby1992; Fodor, Reference Fodor1983; Pinker, Reference Pinker1997), but also appear in more implicit form in cognitive neuropsychology, where modular interpretations are often supported by dissociation logic (Caramazza, Reference Caramazza1986; Coltheart, Reference Coltheart2011). We argue that once metabolism is treated as a globally shared and dynamically allocated constraint, architectures that assume strong functional independence between cognitive subsystems become difficult to justify. Metabolic considerations, therefore, do not merely constrain which computations are feasible; they also place limits on which architectural descriptions remain stable across contexts.

The evaluative framework proposed by Haueis and Colaço emphasizes that metabolism constrains the computational operations biological systems can implement. An important implication concerns the degree of independence cognitive subsystems can plausibly exhibit. Many architectures assume that distinct functions (e.g., attention, memory, and decision-making) can be modeled as relatively autonomous mechanisms, justified on computational or behavioral grounds such as dissociations between tasks or selective impairments. However, if neural systems operate under shared energetic constraints, functional autonomy becomes metabolically suspect.

The issue is not that shared resources eliminate modularity – many biological and engineered systems maintain functional specialization despite common resource pools. Rather, it pertains to whether cognitive subsystems can remain operationally independent under fluctuating energetic demands (Attwell & Laughlin, Reference Attwell and Laughlin2001). Energy in the brain is neither locally unlimited nor strictly compartmentalized: even when global metabolic expenditure remains relatively stable across tasks, local increases in energy demand must be balanced by redistributions across regions or processes (Lennie, Reference Lennie2003; Raichle & Gusnard, Reference Raichle and Gusnard2002; Raichle & Mintun, Reference Raichle and Mintun2006). Although mechanisms such as neurovascular coupling allow for localized regulation of metabolic supply, they operate within a globally constrained system in which increases in demand must ultimately be balanced against overall energetic limits (Friston, Reference Friston2010; Logothetis, Reference Logothetis2008; Raichle, Reference Raichle2015). Under such conditions, sharp and stable boundaries between cognitive components are unlikely to hold across all contexts, and functional decompositions may represent approximate descriptions that hold under stable energetic conditions but degrade when metabolic constraints are perturbed (Moreau, Reference Moreau2025).

Empirical observations from metabolic perturbations illustrate this point. Acute energetic manipulations – such as fasting, fatigue, or sustained cognitive effort – tend to produce broad, graded changes across multiple cognitive domains rather than isolated deficits confined to a single function (Bamberg & Moreau, Reference Bamberg and Moreau2025; Chee & Chuah, Reference Chee and Chuah2008; van der Linden, Frese, & Sonnentag, Reference van der Linden, Frese and Sonnentag2003). Crucially, this is not consistent with the simple domain-general “resource models” that Haueis and Colaço rightly critique for failing to capture the complexity of metabolic regulation. Rather, the implication is architectural: when metabolic resources become scarce or must be redistributed, multiple processes are affected simultaneously, even when they rely on partially distinct neural circuits, suggesting the system is less cleanly partitioned than functional descriptions imply (Anderson, Reference Anderson2010; Sporns & Betzel, Reference Sporns and Betzel2016).

One way to reconcile this with the widespread success of modular descriptions is to treat modular structures as local optima within a high-dimensional constraint space defined by task demands, neural organization, and metabolic efficiency (Moreau & Wiebels, Reference Moreau and Wiebels2022). Under stable energetic conditions, these optima may produce the appearance of well-defined modules, even though the underlying system remains metabolically and dynamically coupled. Perturbations in energetic availability – whether due to physiological states, prolonged cognitive effort, or developmental and aging processes – may shift the system away from these equilibria, revealing underlying couplings that modular descriptions obscure. From this perspective, metabolism does not merely constrain the implementation of computations; it shapes the landscape of possible cognitive architectures, and, like cognitive interventions, can be leveraged to formally test them (Moreau, Reference Moreau2022). Some architectural configurations may be computationally elegant yet metabolically unsustainable, whereas others may endure precisely because they satisfy energetic constraints.

These considerations extend the target article’s framework in two ways. First, metabolic plausibility should be applied not only to the operations posited by cognitive models but also to the architectural assumptions those models rely upon: if a model presupposes strong independence between subsystems, that independence should be assessed under realistic energetic conditions. Second, and more concretely, metabolic perturbations provide a diagnostic tool for testing architectural claims: if two processes are truly modular, transient energetic stress should selectively impair the more metabolically demanding one; if both degrade in tandem, this reveals coupling that contradicts modular assumptions. Leveraging how cognitive performance changes under energetic stress, researchers can thus assess whether proposed decompositions reflect genuine architectural separations or merely stable operating points within a coupled system.