BARANDIARAN; DI PAOLO & ROHDE (2009) – Defining Agency. individuality, normativity, asymmetry and spatio-temporality in action


BARANDIARAN, Xabier. DI PAOLO, Ezequiel. & ROHDE, Marieke. Defining Agency. individuality, normativity, asymmetry and spatio-temporality in action. Journal of Adaptive Behavior,  (Special Issue on Agency), v.10, p.1-13, 2009.


  • We identify three conditions that a system must meet in order to be considered as a genuine agent: a) a system must define its own individuality, b) it must be the active source of activity in its environment (interactional asymmetry) and c) it must regulate this activity in relation to certain norms (normativity).
  • On this basis, we define agency as an autonomous organization that adaptively regulates its coupling with its environment and contributes to sustaining itself as a consequence. We find that spatiality and temporality are the two fundamental domains in which agency spans at different scales. p.1
  • The concept of agency plays a central role in contemporary cognitive science as a conceptual currency across different sub-disciplines (specially in embodied, situated and dynamical approaches—Brooks 1991, Beer 1995, Pfeifer & Scheier 1999). It owes this central role to its capacity to capture the notion of a behaving system while avoiding the endless discussions around alternative foundational terms such as “representations”, “intentions”, “cognitive subject”, “conscious being” or “mind”. While an insect-like robot already seems to be a minimal instance of agency, the concept is open enough to also cover humans or even collective organizations. p.1
  • Despite the difficulty to provide a clear and precise answer to these questions, a loose or metaphorical concept of agency has helped to re-conceptualize cognitive systems as inherently situated while grounding intelligent capacities on behavior-generating mechanisms (as opposed to abstract symbolic algorithms). p.1
  • Russell and Norvig in their classical AI handbook (1995: 33) propose that “an agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through effectors”. Maes (1994), on the other hand defines an agent as “a system that tries to fulfill a set of goals in a complex, dynamic environment”; Beer (1995) considers an agent “any embodied system [that pursues] internal or external goals by its own actions while in continuous long-term interaction with the environment in which it is situated”, while Smithers (1995: 97) states that “agent systems are systems that can initiate, sustain, and maintain an ongoing and continuous interaction with their environment as an essential part of their normal functioning”. After an extensive review of different definitions of agency (including some of those previously mentioned), Franklin and Graesser (1996) conclude that “an autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future”. Kauffman (2000) has defined an agent as a system that “can act on its own behalf in an environment”. Following his work, Ruiz-Mirazo and Moreno defend that minimal autonomous agents are those chemical systems capable of actively constraining their boundary conditions for self-maintenance (Ruiz-Mirazo & Moreno 2000). In a parallel manner, Christensen & Hooker (2000) state that “[a]gents are entities which engage in normatively constrained, goal-directed, interaction with their environment” (p.133). p.2
  • Abstracting away from the particularities of the above definitions we can generalize that agency involves, at least, a system doing something by itself according to certain goals or norms within a specific environment. p.2
  • From this description, three different though interrelated aspects of agency follow immediately: (i) there is a system as a distinguishable entity that is different from its environment, (ii) this system is doing something by itself in that environment and (iii) it does so according to a certain goal or norm. A generative definition of agency has to account, at least, for these three requirements. p.2


  • First of all, in order for a system to be an agent, there must be a distinction between the system and its environment. This we shall call the individuality condition. The identity of an agent as an individual distinguishable from its environment is often taken for granted or seen as trivially irrelevant. Any characterization of agency is then limited to the establishment of the kind of relationship (representational, informational, intentional, adaptive, etc.) between a pre-given “agent” and its world. However, neither a specific environment nor agentive relations with this environment can exist without the constitution of an agent as na individuated system. p.2
  • A concept of agency that cannot account for the way in which an agent defines itself as an individual requires another agent (the observer) to perform the system-environment distinction. If then we have to justify the identity of this observer agent by means of another one and so on, we enter an infinite explanatory regress. In contrast, an entity capable of distinguishing itself as an individual in the absence of an observer, like Jonas proposes for the case of living organisms, does not suffer from this problem (1 – This remark applies to agents once they are in full enjoyment of their agential character. But it does not preclude the possibility that the ontogeny and evolution of different forms of agency is not itself highly dependent on links to a community of other agents and environmental factors. A self-defined identity does not happen in a vacuum and is inevitably tied to a web of necessary relations to develop and survive.) p.3
  • Therefore, the first condition for the appearance of agency is the presence of a system capable of defining its own identity as an individual and thus distinguishing itself from its surroundings; in doing so, it defines an environment in which it carries out its actions. p.3

Interactional Asymmetry

  • Once an individual is in place, exchanges of matter and energy are inevitable at some level; the system is coupled to its environment. However, the concept of agency is intuitively associated with that of action, not mere system-environment coupling or exchange. An agent is a system that does something as opposed to other natural entities to which we attribute no specific actions except metaphorically (e.g., “The sun rises”). In other words, an agent is a source of activity, not merely a passive sufferer of the effects of external forces. Similarly, an agent is not driven to act by internal, sub-systemic modules, which subordinates the system to the triggering or isolated functioning of a local mechanism. In a sense yet to be properly disclosed, an agent as a whole drives itself, breaking the symmetry of its coupling with the environment so as to modulate it from within. We call this condition interactional asymmetry. p.3
  • One way to understand interactional asymmetry in terms of the causal origin of action events is to consider, as others have done, an agent as responsible for managing and gathering the energy resources for action. For this line of thinking, the asymmetry requirement is expressed in terms of the capacity of the system to constrain energy flows to sustain coordinated processes that are in turn re-used by the system in a circular manner (Kauffman 2000, Ruiz-Mirazo & Moreno 2000). p.3
  • However, being a source of activity does not imply trying to constantly avert the effect of environmental forces through the investment of internally channeled energy, but often, on the contrary, being able to “surf” these effects in a specific direction. p.3
  • An agent is a system that systematically and repeatedly modulates its structural coupling with the environment. We therefore define interactional asymmetry as the condition describing a system as capable of engaging in some modulations of the coupling and doing so at certain times, but not necessarily always (and, for extreme cases, just capable of halting a coupling). p.4


  • When considering agency we presuppose that the interaction is not random or arbitrary but makes some “sense” for the agent itself. Agents have goals or norms according to which they are acting, providing a sort of reference condition, so that the interactive modulation is carried out in relation to this condition (2 – We shall use the terms “norm” and “goal” interchangeably. Despite the notion of “norm” is generally applied to a procedure or a limit condition that must be respected whereas that of “goal” refers to specific reference states (get to position X, grasp object Y, attain result Z), for minimal cases both terms might be treated equivalently since both capture a necessary or desired condition that a process must achieve. Explicit distinctions between norms, rules, goals, intentions, desires, plans, etc. would demand reference to more elaborate forms of agency that remain out of the scope of this paper.) p.5
  • We can only make sense of norms as the result of a specific set of conditions that both enable and demand a system to distinguish between different physical outcomes of its coupling with the environment. Normativity is an essential component of agency, even if its presence can be stronger or weaker, as a degree of improvement, of increasing/decreasing adequacy, of gradual functional achievement, etc. This is the case independently of whether norms are linked directly or indirectly to vital requirements (the self-maintenance of the agent’s biological infrastructure) or are acquired and embodied in other self-sustained forms of life (psychological, cultural, etc.). Again, it is insufficient that we, as observers, make judgments on behalf of the agent about the “adequacy” of its behavior in relation to some of our own norms, standards or goals (epistemic, artistic, ethical, functional or otherwise). p.5
  • The first thing to note is that the three requirements are necessary conditions for agency but none of them is sufficient on its own (neither any two of them without the third). Yet, not all of them stand in the same relationship to each other. The individuality condition appears as a precondition for the other two. Neither asymmetry nor normativity would make much sense in the lack of an individualized system to which these properties can be attributed. p.6
  • The picture that comes out of this tradition is that the required minimal living organization is that of a far-from-thermodynamic- equilibrium system, a metabolic network of chemical reactions that produces and repairs itself, including the generation of a membrane that encapsulates the reaction network while actively regulating matter and energy exchanges with the environment. From this point of view, organisms are integrated and active systems that must continuously interact with their environment to self-generate and maintain their own dissipative organization. This minimal (or proto-cellular) living organization comes to capture the essence of life, for even complex multicellular organisms ultimately respond to the same logic of networked self-regeneration and self-regulation through its openness to the environment. These minimal models already provide a first empirically addressable sense of individuality and normativity without having to invoke abstract mentalistic entities such as “propositional beliefs” and “motivations” or without having to reduce the phenomenology of agency to the “selfishness” of a replicating molecule (Dawkins 1976). p.6-7
  • The satisfaction of the individuality condition is almost straightforward: the very organization of a living system is self-asserting, by continuously regenerating itself and its boundary, living systems are demarcating themselves from their surrounding as unified and integrated systems. In doing so they also carve an environment out of an undifferentiated surrounding: the organization of the system (the way in which components processes are nested with each other building up a whole) determines which environmental features are “relevant” to it, i.e., which chemical components in the environment can affect it or are needed for its continued existence. In this way, the environment is not just what lies outside the system as demarcated from the observer’s point of view but is specified by the system through the set of boundary conditions that affect it. p.7
  • In turn, this is where living individuality naturally leads to normativity: component reactions must occur in a certain manner in order for the very system to keep going, environmental conditions are good or bad for the continuation of the system, the system can fail to regain stability after a perturbation, etc. This normative dimension is not arbitrarily imposed from the outside by a designer or external agent that monitors the functioning of the system and judges according to her interests. It is the very organization of the system that defines a set of constraints and boundary conditions under which it can survive (Christensen & Bickhard 2002, Barandiaran 2007, 2008 and Mossio et al. 2009). In this sense, living systems are subject to a permanent precariousness (Di Paolo 2009) that is compensated by its active organization. This precariousness implies that whatever the organism is doing (i.e. whatever its factual functioning is) there is something that it ought to do; not for an external observer but for itself, for the continuation of its very existence.  p.7
  • The permanent need for external matter and energy and the fragility of living systems, sooner or latter, leads to interactional asymmetry: any organism must actively seek for energy gradients and regulate its relation with the environment in order to compensate or avoid potentially destructive perturbations. So, over the most minimal metabolic network endowed with a membrane, even very simple life forms posses adaptive mechanisms that operate detecting and regulating internal and interactive processes. p.7
  • an agent is an autonomous organization capable of adaptively regulating its coupling with the environment according to the norms established by its own viability conditions. p.8
  • It is the deep circularity and entanglement between networked processes, the self-maintaining conditions they generate and the interactions that the system establishes with the environment what makes agents so challenging to model and understand. p.8
  • Our definition of autonomy (much in the line of Varela 1979) can be applied to other domains. For instance, networked interdependent processes can be chemical reactions, molecular structures, physiological structures (like tissues or organs), neurodynamic patterns at the large scale, sensorimotor loops, social habits, etc. This way, agency does not have to be subordinated to biological/metabolic organization but can appear at different scales responding to a variety of autonomous pro-cesses. p.8-9
  • What remains central to our definition is that for any agentive engagement of a system with its environment its identity must be jeopardized at the proper level and that the interaction must involve a process of compensation for deviations from a norm that is generated from within (both, the norm and the compensation). It is in this sense that the interaction becomes meaningful for the agent, that the agent makes sense of a situation: actions are guided by the need to compensate the threatening deviation from a norm and environmental processes are integrated into the interaction as relevant for the achievement of such compensation. We call this process sense-making (Di Paolo, 2005, Thompson 2004) for what would otherwise be a mere event or occurrence becomes valued. The threat must not be interpreted exclusively in terms a direct challenge to the continuation of the agent. It can take the form of a tension or imbalance that, without directly challenging the identity of the system, still provokes an involvement of the whole system in its attempt to counteract the imbalance with the effect that more direct threats are consequently averted. p.9
  • While capturing those features that are essential to minimal forms of agency, the definition remains open to further conditions, interdependencies, hierarchies of modulation, forms of coupling, etc. that might account for more complex types of agency. Similarly, we should not expect natural agents to operate at a single level of organization. Most will involve different scales of autonomy (metabolic, immune, neuro-dynamic, social, etc.) forming nested hierarchies of adaptive regulation (like metabolic monitoring mechanisms modulating behavioral responses or neuro-dynamically induced psychosomatic disorders in the immune system). But leaving aside the sophisticated cases that involve different scales of autonomy it is fundamentally through the spatial and temporal dimensions that agency expands in complexity. p.9
  • Agency is inherently a temporally and spatially extended process. When we say so, we mean not only that the processes described have an essential temporal or spatial extension in the eye of the observer, but also that an agent’s own perspective has temporal and spatial structure and that this depends on its form of agency. p.9
  • The time span of the interdependencies between such processes and their precariousness (their extinction outside the organization that sustains them) is also crucial to understand the self-maintenance of the system and its margin to compensate decay and perturbations. In addition, different rhythms, temporal scales and phenomena of synchronization and co-variation might be found at the core of constitutive processes (Buzsáki 2006). Secondly we also notice that the adaptive modulation of a coupling makes agency unfold temporally: in order for a system to regulate itself there must be some buffering or distance between the immediate perturbation and the possibility of compensating for it. p.9
  • There is also a sense in which spatiality turns out to be relevant for many forms of agency (certainly for living systems), that is, the spatial or topological properties of the processes that constitute the autonomous organization of the system and also its coupling. On the one hand constitutive processes (and interdependent relationships between them) might crucially rely on spatial aspects; for instance the formation of spatially structured patterns in self-organized processes such as convection flows (Hanczyc et al. 2007). p.9
  • A sensorimotor coupling is, primarily, a coupling between a geometrical space and a dynamical system. This implies, first of all, that behavior cannot be taken to be exclusively the result of extracting statistical properties or patterns from a string of predefined sensory inputs and the production of an adequate response output. Situatedness provides much more complex and flexible possibilities for action. p.9
  • Poincaré (1895) has argued that the Euclidean geometrical properties of an agent’s world are due to its sensorimotor situatedness in a spatial environment and to its capacity to enact invariant properties (such as continuity of space, dimensionality or homogeneity) through sensorimotor structuring of its experience (like active visual tracking, reversibility of perceptions and invariance of shape upon movement around an object). Even when he was not directly concerned with the nature of agency Poincaré conceptualized spatial properties as arising from the possibilities and regularities of bodily actions. Motility in a spatial environment equips an agent with the possibility to cope recurrently with the perturbations it encounters and to span them onto a domain of interactions and flexible sensorimotor correlations. With such a way of recurrent modulation, an agent has the possibility to restore situations at will, exploiting the structural invariants of the sensorimotor coupling with the environment that it thus creates. Therefore, the challenge is to reconstruct the spatio-temporal dimensions of the environment of the agent not from the point of view of the observer scientist or the modeler, but from the frame of reference of the agent itself. p.10
  • Some insects, mammals and birds clearly exploit not just the orderly, but also the metric properties of their couplings with the environment. Their rich sensorimotor inventory, afforded by the nervous system’s fast and flexible way of linking sensors and actuators, allows them to further increase the degree of mediacy between the surface effect of the stimulus and its meaning for the system by adding another layer of abstraction to its perspective on the spatio-temporality of its coupling with the environment. This transition in spatio-temporality coincides with a transition in agency. p.10
  • The reality of our embodied behavior shows, by contrast, that our interactions with the world in the vertical dimension are strongly influenced by the vestibular sense (due to gravity), which makes them very different from our interaction with the world in the horizontal plane (e.g., Gibson 1952). Similarly, we make an explicit spatial analogy of time as an arrow in thinking and language (cf. Lakoff & Johnson 2003, Rohde 2008). Such symbolic spatio-temporality, that lumps together a diverse set of sensorimotor couplings with the world pushes the stimulus and its meaning even further apart. p.10
  • Other questions have to do with the relationship of co-dependence between system and environment. Although a first approximation to the problem required distinguishing the system from its environment, agency (especially when considering recurrent sensorimotor situatedness) leads to a deep entanglement of an agent with its environment. Yet, despite its “being-in-theworld” an agent does selectively couple with environmental features asymmetrically integrating them on its behavioral organization. A number of questions follow: How does niche construction (for example) relate to agency? Should those environmental features that recurrently depend on the agent be considered as part of the agent? What is the status of tools as mediators between agents and environments? p.11
  • And yet, despite the fact that our definition is, admittedly, not yet complete there are concrete and practical consequences that can be extracted for the study of adaptive behavior: a) mere sensorimotor coupling on its own is too weak a condition for agency, modulation of interactions need also be considered; b) systems that only satisfy constraints or norms imposed from outside (e.g. optimization according to an externally fixed function) should not be treated as models of agency; and, c) the identity of an agent cannot be divorced from its behavior, therefore, some kind of feedback between the agent’s behavior and the selfmaintenance of its organization should be included in our models (i.e. the agent must “benefit” or “suffer” the consequences of its action in a manner that is relevant for its continued activity). Finally, it must be stressed that models of agency can explore different aspects of our definition without the system fully satisfying the three requirements. p.11
  • The adaptive regulation of behavior needs not be exclusively subordinated to the viability constraints imposed from biological “survival conditions”. Instead, it can be equally governed by the need to maintain neuro-dynamic and behavioral organization in terms of self-maintenance of habits, coherence of behavior, “psycho-dynamic” stability, etc. (Di Paolo 2003, Barandiaran & Moreno 2006, Barandiaran & Di Paolo 2008). p.12


Artigo na íntegra


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s