Agency is a foundational yet contested concept across various subfields of philosophy, biology, cognitive science and psychology. Often, it is treated as a philosophical placeholder or intuition about whatever distinguishes purposeful behaviour from mere dynamics and passivity. However, it also operates with increasing proximity to computational approaches and methods that aim to model or approximate the continual, open-ended regulation of action that is characteristic of biological agency in dynamic environments. In my thesis, drawing on the framework of reinforcement learning (RL) and comparative methods, I attempt to test the boundaries of RL style accounts of agency, evaluating its use as a cross-species and cross-systems framework that is sensitive to different forms of control and constraint. I then turn to a more critical analysis of current RL-based approaches to agency, aiming to clarify when its models confer genuine explanations of biological behaviour, andwhen they function more as simplified explanatory strategies or otherwise convenient fictions.

At a high-level, RL refers to the process by which an agent learns to act through feedback, using reinforcers that signal whether outcomes are better or worse. Within the RL paradigm, there is often convergence around three core assumptions: (i) behaviour is organised around the maximisation of cumulative reward (i.e., a scalar quantity), (ii) objectives are explicit and well-defined, and (iii) control is mediated by environmental feedback (i.e., control as closed-loop adjustment). This template has helped to uncover a wide range of biological phenomena, from dopaminergic prediction errors to developmental shifts in exploration strategies, and appears to provide a blueprint for designing artificial agents capable of surpassing human performance in specific (well structured) tasks and environments. More recently, such accounts have been extended to the concept of agency, which is approximated as the ability of a system to regulate its interaction with the world so as to preserve and extend its future possibilities - a kind of “control by consequences”, where behaviour is shaped by how well things go, and the shaping mechanism is explicit.

Yet, the same simplifying assumptions and parsimony that give RL its operational edge also prompt questions about its ecological validity as a framework for thinking about agency. Whilst RLs core assumptions are plausible in restricted settings, its models abstract away critical features of the messy realities of biological life, and the unique ways in which agents have evolved to be sensitive to their environmental niche. Living systems, from the seemingly “simple” to the cognitively “complex”, do not simply optimise external payoffs, but can invent and revise goals and norms, and act under evolved constraints. Biological agents exhibit actions that are driven by intrinsic motivations and homeostatic needs, which, despite being approximated through various heuristics and biases, resist straightforward reduction to externally specified rewards or computations. A closer look at these asymmetries, particularly when dynamics are tied to evolutionary and developmental processes and the organisational roots of motivated behaviour in regulatory control, reveals the current scope of the RL paradigm. Its formalisms appear well suited to modelling some features of high-level human behaviour, yet unable to capture the fidelity (and generative complexity) of motivated behaviour in biological agents and their living ecologies.

A more naturalistic account requires bringing control and constraint into focus as an organising principle of agency. Here, agential behaviour is modulated by multiple, overlapping constraints, including metabolic budgets, developmental trajectories, ecological affordances, and evolutionary histories. These constraints are organised in ways that notonly determine how agents learn and decide, but also what they value and how those values change, what counts as a goal, how goals are revised, and how trade-offs are tolerated or managed. They are layered and interdependent, with physiology gating developmental plasticity, development conditioning ecological engagement, and ecology canalising evolutionary strategies. To the extent that agency can be contextually understood as the product of layered physiological, developmental, ecological, and evolutionary constraints, RL models can provide naturalistic explanations of agent-like behaviour only insofar as its abstractions align with, or approximate, a given constraint architecture. The aim of the project is thus to develop a constraint-sensitive account of agency that clarifies the operational boundaries of RL. More generally, I aim to show how these forms of control and constraint organise the functions of agency across scales and substrates, and the utility of RL as both a computational paradigm and a lens that can be deconstructed and reassembled to better understand the comparative realities of agenthood in biological systems. The question is therefore not whether RL “gets agency right”, but under what conditions the abstractions of RL can track the biological dynamics of agenthood, and when alternative frameworks are needed.