1.1 Acting for Reasons

From Aristotle onwards, most philosophical work on agency has focused on human intentional agency. In contemporary analytic philosophy, Anscombe (Reference Anscombe1957) and Davidson (Reference Davidson1963) have been very influential in promoting the view that agency is best understood as a capacity for intentional action. Both agreed that (a) to act intentionally is to act for a reason and that (b) the notion of intentional action is more fundamental than the notion of action.

The following often quoted passage from Anscombe’s Intention (Reference Anscombe1957) illustrates the first claim:

Intentional action is thus behavior whose correct explanation cannot be purely mechanistic but requires us to make reference to the reason the agent has for so acting, a reason that identifies the goal or aim of the action and thus makes it intelligible in the agent’s eyes. For instance, if you ask a pedestrian why they are crossing a street and they answer that they are on their way to the shops on the other side, the pedestrian has given you a reason for their action. The reason given here may be taken as shorthand for a more complete reason explanation that takes the form of a practical syllogism with a desire and a belief as premises and the action they rationalize as the conclusion. In Davidson’s own words:

In our example, (a) would be the desire to go to the shops and (b) the belief that crossing the street is a way to go to the shops.

The second point of agreement between Anscombe and Davidson is that the notion of intentional action is more fundamental that the notion of action, or, in other words, that the latter is derivative from the former. Let us go back to our pedestrian attempting to cross the street and imagine this time that as he is doing so, he jumps backwards and bumps into the woman walking behind him. Asked why he jumped backward he may answer that it was to avoid being hit by a car that had suddenly turned the corner at high speed. Asked why he bumped into the woman behind him, he might answer instead that this action was involuntary on unintentional. According to Anscombe and Davidson, one and the same event is susceptible of a number of different descriptions. Described as jumping backward, the event is an intentional action insofar as the pedestrian had a reason to do so. Under the description bumping into the woman behind him, the event is not an intentional action, yet it qualifies as an action insofar as this description refers to the same event that can also be described intentionally. We can contrast this with the following scenario: Our pedestrian is once again attempting to cross the street but this time he is the victim of a dizzy spell, loses his balance, falls backwards, and bumps into the woman behind him. This time, his falling backward and his bumping into the woman cannot be given intentional descriptions; rather they are mere happenings.

So far, we have seen that Anscombe and Davidson agree on the fundamental character of the notion of intentional action and agree that an action is intentional to the extent that it is explainable by reasons the agent has for so acting. But we still have to answer a third question, namely what is the nature of this explanation relation between reasons and actions. This is where Anscombe and Davidson clearly part ways.

In the 1950s and early 1960s, an important debate was sparked on whether agents’ reasons for their actions were also the causes of these actions. Influenced by Wittgenstein (1952), some philosophers (e.g., Melden, Reference Melden1961; Kenny, Reference Kenny1963; R. Taylor, Reference Taylor1966) argued that the connections between reasons and actions were logical, conceptual, and normative and as such could not be causal. Others (C. Taylor, Reference Taylor1964; Wilson, Reference Wilson1980) argued that explanations of actions are teleological explanations – in other words, explanations in terms of goals – and as such not analyzable as causal explanations. Anscombe herself considered that the relation between reasons and action was best approached in epistemological terms and involved a special form of practical knowledge. In contrast, Davidson (Reference Davidson1963) argued that reason explanations are causal explanations and did much to rebut the anti-causalist arguments that purported to show that reasons couldn’t be causes. Under his impetus, the causalist approach was revived and refined and became, in the eyes of many, the dominant position in philosophical action theory.

I will come back in Section 3 to the notion of practical knowledge central to Anscombe’s theory of action. For now, I shall concentrate on causal theories of action (CTA, for short), their evolution and variants, and the limitations and shortcomings they face.

1.2 Varieties of Causal Theories of Action (CTA)

Broadly speaking, causal theories consider that action is behavior that can be characterized in terms of a certain sort of psychological causes. Yet, versions of the causal approach can take widely different forms depending on (a) what they take the elements of the action-relevant causal sequence to be and (b) what part of the sequence they identify as the action.

On the basis of the second criterion, one can distinguish three broad types of causal theories. On one view, one should characterize actions in terms of their causal power to bring about certain effects, typically bodily movements and their consequences. Accordingly, proponents of this view will tend to identify an action with mental events belonging to the earlier part of a causal sequence, such as tryings (Hornsby, Reference Hornsby1980).

Conversely, one may hold that what distinguishes actions from other kinds of happenings is the nature of their causal antecedents. Actions will then be taken to be events (typically again, bodily movements and their consequences) with a distinctive mental cause. This second type of causal theory was made popular most notably by Davidson (Reference Davidson1963) and Goldman (Reference Goldman1970). When the label “causal theory of action” is used in its narrower sense, it usually refers to approaches of this second type.

A third possibility is to consider actions as causal processes rather than just causes or effects and to identify them with, if not the entire causal sequence, at least a large chunk of it. On this third view, actions are characterized, one may say, in terms of their distinctive causal structure, rather than relationally in terms of their causes or effects.

Each of these three types of theories can be made to look more or less plausible depending on what the elements of the causal chain are taken to be and on how they are characterized. For instance, some theories countenance only beliefs and desires, while others view intentions, volitions or tryings as essential elements of the action-relevant causal sequence.

1.3 Belief-Desire Versions of CTA and Their Limitations

The earlier belief-desire versions of the causal theory, made popular most notably by Davidson (Reference Davidson1963) and Goldman (Reference Goldman1970), claimed that what distinguishes an action from a mere happening is the nature of its causal antecedent, conceived as a complex of some of the agent’s beliefs and desires. This complex was held to both rationalize the action and cause it. According to Davidson (Reference Davidson1963), for instance, the causal antecedent of an action, what he calls a primary reason, is a combination of a pro-attitude toward actions of a certain kind and a belief that this action is of that kind. Pro-attitudes include “desires, wantings, urges, promptings, and a great variety of moral views, aesthetic principles, economic prejudices, social conventions, and public and private goals and values insofar as these can be interpreted as attitudes of an agent directed toward actions of a certain kind” (1963: 4).

The elegant simplicity of the belief-desire theory may have contributed to the attraction it has exerted. First, it brings into line the justificatory and the explanatory role of reasons by insisting that in cases where reasons genuinely explain, the reason-providing intentional states cause the actions for which they provide reasons. For instance, I might have two reasons to go to my office this morning: I have an appointment with a student and a technician is coming to fix the heating system. Yet, if I have forgotten about the technician, the only reason that explains my going to my office is my appointment with the student. The theory thus fosters the hope of narrowing the gap between the normative realm of explanations by reasons and the natural realm of causal explanation. A second attraction of the theory is its ontological parsimony. On this view, to say that somebody acted with a certain intention is just to say that his actions stood in the appropriate relations to his desires and beliefs. No distinct state of intending nor any special type of mental event such as willings, volitions, acts of will, settings of oneself to act, or tryings are postulated and thus, no embarrassing entity is added to the world’s furniture.

However, it soon appeared that simple belief-desire versions of the causal theory were faced with several difficulties. These difficulties make it doubtful whether it is either a necessary or a sufficient condition for an event to qualify as an action that it have as an antecedent a belief-desire complex. They also highlight important limitations in what the belief-desire approach can explain.

1.4 Intentions as Distinctive States

Acknowledging the existence of future-directed intentions forces one to admit that intentions can be states separate from the intended actions or from the reasons that prompted the action, and, as Davidson (Reference Davidson1978) notes, once this is admitted, there seems to be no reason not to allow that intentions of the same kind are also present in all or at least most cases of intentional action. Once intentions are acknowledged as separate mental states and not just relational constructs, a new issue arises. Can these states be given a reductive analysis? In other words, can they be assimilated to complexes of desires and beliefs, to special kinds of beliefs or judgments, or do they form a sui generis and irreducible class of mental states?

Davidson himself (Reference Davidson1978) saw that the commitment-to-action feature of states of intending created a problem for the kind of reductive analysis that took intentions to be reducible to combinations of beliefs and desires. He argued that intention should be seen as a special kind of evaluation of conduct. In his view, although both desires to act and intentions are evaluative judgments, they are different kinds of judgments. Davidson distinguishes between prima facie judgments, all things considered judgments, and all-out judgments. Prima facie judgments are of the form “A is better than B in light of consideration C.” All things considered judgments constitute a subclass of prima facie judgments that assess the relative merits of potential courses of action in light of all the considerations deemed relevant by the agent. “A is the best course of action in the light of all relevant considerations” is an all things considered judgment. Both prima facie and all things considered judgments are conditional or relational judgments. In contrast, all-out judgments are absolute or unconditional judgments of the form “A is the best course of action.” According to Davidson, the transition from an all things considered judgment to an all-out judgment is governed by a substantial principle of practical rationality, which he calls “the principle of continence.” This principle enjoins agents to perform the action judged best on the basis of all available reasons and it is such all-out judgments that bridge between practical reasoning and intentional actions. Desires to act correspond to what he calls prima facie judgments, judgments that actions of a certain kind are desirable insofar in the light of certain considerations. By contrast, an intention to act corresponds to what Davidson calls an all-out judgment. In making an all-out judgment as opposed to a prima facie judgment, we settle on a course of action. Intentions are thus associated with actions in a way that mere desires are not. By acknowledging the existence of intentions as separate states, Davidson makes a fair attempt at coming to grip with the commitment-to-action feature that appears to be characteristic of intentions as opposed to mere desires. By analyzing intentions as a special kind of evaluative judgments, all-out judgments, he avoids having to postulate a sui generis kind of mental entity. Intentions, together with other pro-attitudes, are deemed to belong to the general class of evaluative judgments.

Despite its merits, Davidson’s revised position fails to capture the full import of the commitment-to-action-feature of intentions. As argued by Bratman (Reference Bratman1987), two dimensions of this commitment should be distinguished. The first dimension, what he calls the volitional dimension, can be characterized by saying that “Intentions are, whereas ordinary desires are not, conduct-controlling pro-attitudes. Ordinary desires, in contrast, are merely potential influencers of action” (1987: 16). There is yet, according to Bratman, a second dimension of commitment, what he calls the reasoning-centered dimension of commitment. What is at stake here are the roles played by intentions in the period between their initial formation and their eventual execution. First, intentions have what Bratman calls a characteristic stability or inertia: Once we have formed an intention to A, we will not normally continue to deliberate whether to A or not; in the absence of relevant new information, the intention will resist reconsideration, we will see the matter as settled and continue so to intend until the time of action. Second, during this period between the formation of an intention and action, we will frequently reason from such an intention to further intentions, reasoning from instance from intended ends to intended means or preliminary steps. And third, this intention will constrain the other intentions one may form. For instance, if I intend to go see a movie tonight, I cannot consistently intend to go to a concert at the same time.

One may think that Davidson’s analysis of intentions as all-out judgments is meant to capture the volitional dimension of commitment and that it may perhaps as well account for the stability of intentions once formed. Yet the other aspects of the reasoning-centered dimension of commitment seem to fall beyond its scope. As Bratman suggests, rather than attempting to give reductive analyses of intentions, it may be more illuminating to take seriously the idea that intentions are distinctive states of mind and that they should be characterized in terms of their own complex network of dispositions and functional roles. As another way to put it, Davidson’s approach is too exclusively backward-looking. He focuses on how intentions are arrived at and in so doing fails to appreciate important aspects of their roles.

Advocates of intentions as distinctive states tend to emphasize a number of functions plausibly attributed to intentions. They argue that intentions have their own complex and distinctive functional role that warrants considering them as forming an irreducible kind of psychological state, on a par with beliefs and desires. As we have just seen in our brief discussion of his approach, Bratman stresses three functions of intentions. First, intentions are terminators of practical reasoning in the sense that once we have formed an intention to A, we will not normally continue to deliberate whether to A or not. In the absence of relevant new information, the intention will resist reconsideration. Second, intentions are also prompters of practical reasoning, where practical reasoning is not concerned this time with whether or not one should A, but is reasoning about means of A-ing. This function of intentions thus involves devising specific plans for A-ing. Third, intentions also have a coordinative function and serve to coordinate the activities of the agent over time and to coordinate them with the activities of other agents.

Philosophers also typically point out further motivational and cognitive functions of intentions (e.g., Brand, Reference Brand1984; Mele, Reference Mele1992). Intentions are also responsible for triggering or initiating the intended action (initiating function) and for sustaining it until completion (sustaining function). An intention to A incorporates a plan for A-ing, a representation or set of representations specifying the goal of the action and how it is to be arrived at. It is this component of the intention that is relevant to its guiding function. Finally, intentions have also been assigned a monitoring function, involving a capacity to monitor progress toward the goal and to detect and correct deviations from the course of action as laid out in the guiding representation. We may call the initiating and sustaining functions motivational functions and the guiding and monitoring functions control functions. Although there seems to be some consensus as to what the initiating and sustaining functions amount to, the situation is much less clear where the control functions are concerned.

The first three functions of intentions stressed by Bratman – their roles as terminators of practical reasoning about ends, as prompters of practical reasoning about means and as coordinators – are typically played by intentions in the period between their initial formation and the initiation of the action. By contrast, the last four functions (initiating, sustaining, guiding, and monitoring) are played in the period between the initiation of the action and its completion. Attention to these differences has led a number of philosophers to develop dual-intention theories of action. For instance, Searle (Reference Searle1983) distinguishes between prior intentions and intentions-in-action, Bratman (Reference Bratman1987) between future-directed and present-directed intentions, and Mele (Reference Mele1992) between distal and proximal intentions. In all cases, an intention of the former type will only eventuate into action by first yielding an intention of the latter type. For the sake of brevity, in what follows, I shall use Mele’s terminology of distal and proximal intentions.

Dual-intention theories make available new strategies for dealing with the difficulties listed earlier. To begin with, they seem to provide at least a partial answer to the problem of causal deviance. First, they suggest that for intentions to cause actions is the right way for them to count as intentional, the intended effect must be brought about in the way specified by the plan component of the intention. In cases of consequential deviance, this condition is not satisfied. For instance, it was not part of Carl’s plan to kill his uncle by hitting him with his car. Second, for intentions to cause behavior in the right way for it to count as an action, it must also be the case that the causal chain linking the distal intention to the resultant bodily behavior include relevant proximal intentions. Cases of antecedential deviance do not satisfy this condition. For instance, in the marriage proposal example introduced earlier, the young man had formed the distal intention to intend to get down on his knees to propose marriage; yet he did not get a chance to form the corresponding proximal intention before sinking to his knees.

Dual-intention theories also open up new prospects toward a solution to the problem of spontaneous and arational actions. According to dual-intention theories, all actions have proximal intentions, but they need not always be preceded by distal intentions. For instance, when, reflecting on a philosophical problem, I start pacing about the room, I do not first engage in a deliberative process that concludes with a distal intention to pace; rather my pacing is initiated and guided by a proximal intention formed on the spot. Automatic, spontaneous, or impulsive actions may then be said to be those actions that are caused by proximal intentions but are not planned ahead of time.

Whether distinguishing between distal and proximal intentions also helps with the problems of bodily movement and action awareness is more controversial and depends on what different authors take the exact contents and functions of proximal intentions (aka, present-directed intentions or intentions-in-action) to be. There are disagreements among dual-intention theorists regarding how much of what is going on representationally in the preparation and execution of an action should be included in the content of intentions.

However, in many versions of dual-intention theories, proximal intentions are conceived as having the function of guiding and controlling the course of an ongoing action. Dual-intention theories thus seem in a position to provide some explanation of the distinctive character of our awareness of our own actions. Insofar as a proximal immediate intention has a sustaining function and is involved in the guiding and monitoring of the action, it does not terminate with the onset of action, but continues as long as the guiding and monitoring continue. As a consequence, it seems possible to reconcile the view that the main difference between actions and simple events lies in their causes with the idea that we are immediately aware that we are acting and that this awareness has a non-perceptual source. Dual-intention theorists may claim that this immediate awareness is in fact an awareness of the content of the proximal intention that is sustaining the action as it unfolds.

Some versions of the dual-intention approach may also have the resources needed to address the problem of bodily movements. For instance, Searle takes the content of intentions-in-action to differ from the content of prior intentions in more respects than just tense. In his view, a prior intention is an intention that is formed prior to the action and that represents and causes it as a whole. In contrast, an intention-in-action is the mental component of the action itself; it presents, causes, and is contemporaneous with, bodily movements. The intention-in-action together with the bodily movements it causes constitute the action. Thus, while the content of my prior intention may simply be that I raise my arm at some point in the future, the content of an intention-in-action includes a detailed representation of the movement of my arm, its speed and trajectory. Insofar as intentions-in-action include detailed representations pertaining to movement, Searle’s version of the dual-intention is in a position to address the problem of the wrong movement. Whether this is also the case for other versions of dual-intention theories depends in part on how they construe the contents of proximal intentions.

1.5 A Dynamic Theory of Intentions: The DPM Model

As we have just seen, dual-intention theories offer ways of addressing at least some of the difficulties that plague early belief-desire versions of CTA. Yet, they typically do not say much about how action guidance and control operate and do not always clearly characterize the different forms of action guidance and control associated with different types intentions. In addition, different versions of dual-intention theories offer divergent views of the contents of proximal intentions and of their exact functions.

I have proposed (Pacherie, Reference Pacherie2008; Mylopoulos & Pacherie, Reference Pacherie, Newen, de Bruin and Gallagher2018) a dynamic hierarchical model of intentions that brings together philosophical work on intentions and empirical work on motor representations and motor control. This model distinguishes three main stages in the process of action specification, each corresponding to a different level of intention and each level of intention having a distinctive role to play in the guidance and monitoring of the action. It thus proposes a threefold distinction among distal intentions (D-intentions), proximal intentions (P-intentions), and motor intentions (M-intentions), hence the DPM model for short. The DPM model specifies the representational and functional profiles of each type of intention, as well as their local and global dynamics, and the ways in which they interact. One of its core tenets is that action control is the result of integrated, coordinated activity across these levels of intention.

In the DPM model, the distinction between D-intentions, P-intentions, and M-intentions is motivated by an analysis of their different and complementary functional roles, of the different types of contents they involve and of their respective temporal scales. D-intentions are in certain respects close kins to the distal intentions of dual-intention theories. D-intentions also function as terminators of practical reasoning about ends, as prompters of practical reasoning about means and plans, and as intra- and interpersonal coordinators. As such, they are subject to norms of practical rationality. An agent’s intentions should be mutually consistent and they should be consistent with all of the agent’s beliefs (strong consistency requirement). The practical rationality constraints that bear on them require the presence of a network of inferential relations among intentions, beliefs, and desires to insure internal, external and global consistency. It is generally admitted that concepts are the inferentially relevant constituents of intentional states and that their sharing a common conceptual representational format is what makes possible a form of global consistency, at the personal level, of our desires, beliefs, intentions and other propositional attitudes. If we accept this common view, it follows that in order to satisfy the rationality constraints they are subject to, D-intentions must have a propositional format and conceptual content. The main difference with earlier conceptions of distal intentions is that the DPM model assigns a specific control function to D-intentions. They are in charge of the rational control of the action. Their role is to ensure that the way the action unfolds – in particular, adjustments of the action plan in the face of unanticipated side effects or difficulties – doesn’t flout the reasons the agent had for their action in the first place or violate norms of practical rationality.

A P-intention often, though not always, inherits an action plan from a D-intention. One of its functions is to anchor this plan in the situation of action. The temporal anchoring, the decision to start acting now is but one aspect of this process. Once the agent has established a perceptual information-link to the situation of action, she must generate a more detailed representation of the action that fits the specification inherited from the D-intention while anchoring it to the situation at hand. The formation of a P-intention depends on the integration of conceptual information about the intended action inherited from the D-intention with perceptual information about the current situation and memory information about actions in one’s motor repertoire to yield a more definite representation of the action to be performed. For instance, whether I use my right or my left hand to reach for a bottle of water may depend on the position of the bottle relative to me or on the presence of obstacles on the path. Once formed, P-intentions are in charge of the situational monitoring and control of the action as it unfolds. This control is more local than the rational control exerted by D-intentions and is concerned with the immediate goal and the situation as currently perceived.

The main novelty of the DPM model is perhaps the introduction of M-intentions. M-intentions involve what neuroscientists call motor representations. Motor representations are not just representations of movements. They encode action goals together with the motoric means for achieving them and do so in a motoric format directly suitable to action execution. The way they encode movements is normally sensitive not just to the immediate motor goal but to the hierarchy of motor goals in which they are embedded. For instance, the same cup will be grasped in different ways depending on whether one wants to bring it to one’s lips or to turn it upside down. Thus, motor representations appear to be themselves hierarchically organized, with a decomposition of more general into more specific motor goals (Jeannerod, Reference Jeannerod2006; Van Elk et al., Reference van Elk, van Schie and Bekkering2014). According to the DPM Model, M-intentions are also responsible for a distinctive form of monitoring and control of an ongoing action. They are responsible for the precision and smoothness of action execution, and motor control typically operates automatically and at a very fine time-scale.

The introduction of M-intentions is motivated in part by the need to distinguish this automatic, unconscious and very fast form of action control from the typically conscious and slower forms of action control exerted at the level of P-intentions and D-intentions. It is also motivated by the failure of dual-intention accounts to offer more than partial solutions to the problem of antecedent causal deviance and the problem of the wrong movement. For instance, consider the marriage proposal example again. Imagine this time that the young man, having already formed the distal intention to get down on his knees to propose marriage, forms the proximal intention to do so upon entering the room where the woman of his heart finds herself. As he approaches her, their eyes meet and he is so overcome by emotion that he suddenly feels weak and sinks to his knees. In this variation on the marriage proposal example, both distal and proximal intentions are present, but the young man’s sinking on his knees was not caused by a motor intention to do so. Thus, causing behavior in the right way for it to count as an action may be a matter of intentions causing behavior via the instantiation of appropriate motor intentions. Similarly, to explain why Brutus failed to stab Caesar in the chest, we may have to say that that there was something wrong with the content of his motor intentions.

Why talk of M-intentions rather than simply motor representations? M-intentions have the functional role of bringing about the actions they represent, whereas motor representations need not have such a functional role. For instance, motor representations are also formed when we observe someone else acting or when we imagine acting (Jeannerod, Reference Jeannerod1995; Filimon et al., Reference Filimon, Nelson, Hagler and Sereno2007). Thus, talk of M-intentions emphasizes the role of such states in action production and the fact that M-intentions share certain core functional characteristics of D-intentions and P-intentions. It may still be objected that M-intentions fail to meet three constraints that genuine intentions should satisfy, namely conscious accessibility, integration with other propositional attitudes and strong consistency, where conscious accessibility and integration are taken as preconditions of strong consistency (Brozzo, Reference Brozzo2017). M-intentions’ failure to meet these constraints – insofar as their contents are not always consciously accessible and their integration with other propositional attitudes is limited – would be problematic if it led to systematic inconsistencies between an agent’s D-intentions or P-intentions and the behavior for which M-intentions are immediately responsible. However, the architecture of the DPM model precludes such systematic inconsistencies. On the one hand, what M-intentions are formed and what contents they have is constrained, although not fully determined, by P-intentions and their contents, so that initially at least M-intentions will not be inconsistent with higher-level intentions. On the other hand, the existence of three levels of monitoring and control normally ensures that emerging discrepancies during action execution are detected and reduced. Thus, while M-intentions may not directly meet strong consistency constraints, the architecture in which they are embedded ensures that they meet them indirectly.

The DPM model proposes to explicate action production and control in terms of a dynamic hierarchy of intentions. As we shall see, the model also provides a useful framework for thinking about the phenomenology of agency (Section 3) and about joint action (Section 4). It is, however, like the other versions of CTA discussed so far, mainly aimed at explaining human intentional agency and as such cognitively demanding. The models discussed so far, including the DPM model, thus appear ill-suited to account for minimal forms of agency of the kind exhibited by creatures and possibly artifacts with more modest cognitive endowments. This is why in the next section, I change tack, attempting to identify minimal forms of agency and their basic building blocks and to understand how they relate to the more sophisticated forms of agency considered so far.

2.1 Representations Revisited

Let me start with two preliminary remarks. First, non-human agents do not form a homogeneous category. It is more than doubtful that all non-human animals, from chimpanzees to earthworms, not to mention paramecia and other unicellular organisms, have the same agential capacities and are capable of a similar range of purposive activities. Thus, we should not expect a one-size-fits-all story about (non-intentional) purposive agency. Second, humans are part of the animal realm. On the one hand, young children have yet to develop some of the capacities that will allow them to become full-blown intentional agents. On the other, as we already discussed in Section 1, many of the actions performed by human adults – habitual, impulsive, spontaneous, or arational actions – may fail to meet the stringent criteria imposed by CTA. Thus, while humans might have or develop agential capacities non-human agents lack, there is no radical discontinuity between human and non-human agency. Indeed, the more basic agential capacities shared by humans and (some) animal species may serve as a necessary scaffold for full-blown human intentional agency.

There are two main routes one may take to account for the purposiveness of non-human agency. The first is representational and aims at showing that one can ascribe belief-like, desire-like, or intention-like states to non-human agents by appealing to more basic mental representations than those needed for full-blown beliefs, desires, and intentions. The second route is more radical and contends that an explication of non-human forms of agency can dispense with representations altogether. The first route is thus deflationary in matters of representational resources, the second is eliminativist. Let us explore the first route first.

2.2 Purposiveness Revisited

In the absence of mentality, consciousness, or representation, one may wonder what makes animal behavior purposive. According to Burge, the behavior of an organism will count as an action if (a) it is endogenously driven, rather than the organism being merely pushed around by its environment, (b) it is coordinated behavior by the whole organism, issuing from its central behavioral capacities, not purely from subsystems, and (c) it fulfills whole animal biological functions, such as eating, navigating, mating, or parenting, that contribute to the fitness, survival or reproductive success of the individual. Thus, on Burge’s understanding of primitive agency, not only snails but even ticks or paramecia should be considered capable of agency. As we will see in the next section, some philosophers would object to a conception of agency that allows for the attribution of agency to paramecia as too thin and as neglecting an important dimension of agency.

First, though, let us consider goal-directedness and its gradations. Several issues immediately arise: Where do the goals come from? Whose goals are they? Do they need to be represented or not? What difference does it make whether a goal is represented or not?

As we just saw, Burge contends that goals (or, as he calls them, targets, since he reserves the term “goal” for targets that are represented by the agent) need not be represented. Rather to count as goal-directed (or, as Burge would prefer, target-directed) behavior must serve a biological function. The relevant biological functions must be biological functions of individuals or whole organisms rather than of subsystems. As Burge puts it: “The spider ingests; its stomach digests” (2009: 256). Finally, for every biological function there is a biological norm that sets a standard or level of performance that is adequate to fulfill the function in question, where the applicability of these norms is independent of any individual’s accepting or appreciating them.

But, then, how do we determine whether a system or subsystem of an organism has a biological function and which? Here, Burge adopts the evolutionary etiology approach to biological functions (e.g., Millikan, Reference Millikan1984; Neander, Reference Neander1991). The idea in a nutshell is as follows. If in the ancestors of the organism, the behavior produced by a system or subsystem contributed to their fitness and reproductive success, then producing this behavior became the biological function of the system in question. The norm associated with a function corresponds to the level of performance adequate to fulfill this function. For instance, hearts have a biological function with an associated biological norm because in evolutionary history pumping blood benefited the organism’s fitness and survival, and organisms with hearts that did a better job at pumping blood had better reproductive success than those with hearts that did a poorer job.

As Van Hateren (Reference van Hateren and Ferrero2022) points out, we can offer a similar story for agency itself. Agency can be regarded as a biological trait that has a function for an organism insofar as it increased the chances that the organism’s ancestors survived and reproduced in the past:

Critics of the evolutionary etiology approach have highlighted several difficulties it confronts. One difficulty is that biologists who study the form and function of organismal design recognize that it is virtually impossible to determine the past action of selection on any particular structure retrospectively and thus virtually impossible to identify its exact biological function (Amundson & Lauder, Reference Amundson and Lauder1994). Second, this approach looks at the past history of systems to identify the goal-directedness of biological functions in a species. But this leaves open a gap between an explanation of the behavior of past members of a species and an explanation of the behavior its current members. The environment may have changed and features of past environments that contributed to explaining the beneficial aspects of the behavior may no longer obtain. For instance, our own ancestors’ great appetite for sugar and fat may have contributed to their fitness in environments where these resources were scarce and it therefore made sense to gorge yourself when you could. Nowadays, though, in parts of the world where such food resources are plentiful and easily accessible, this appetite is often detrimental to our health rather than being conducive to our fitness and survival.

Self-organization theories of agency inspired by the work of Varela offer an alternative account of goal-directedness and norms, that does not rely on evolutionary considerations. For instance, Barandiaran and colleagues (Barandiaran et al., Reference Barandiaran, Di Paolo and Rohde2009) identify three conditions that a system must meet in order to be considered as a genuine agent. As a first condition, a system must define its own individuality, that is be capable of distinguishing itself from its surroundings. Individuality in this sense does not require representations or consciousness. For instance, the immune system of an organism meets the individuality condition insofar as it distinguishes between the organism’s own tissues and foreign bodies (viruses, bacteria, parasites, etc.) that endanger the organism and attacks only the latter. A second condition is that it must be the active source of activity in its environment, what Barandarian and colleagues call interactional asymmetry. In other words, what the system does is neither driven by external forces nor by subsystems within it; rather what it does is driven by the system as a whole. Finally, a third condition is that the system must regulate its activity in relation to certain goals or norms.Footnote ¹ The most fundamental of these goals or norms concern self-maintenance: The main norm or goal governing the activity of an agent is the continuation of its own existence. More specific norms and goals relate to the different ways in which a change in the system’s processes or in the environment can threaten its viability conditions and thus the different ways in which the system can restore its organization as a self-maintaining system. This approach of course allows that more complex systems may have more complex interactions with their environment and indeed more complex environments they interact with (from chemical and physical to social and cultural environments) and, as a result, exhibit richer and more sophisticated forms of agency. Yet, it contends that even proto-cellular systems can be counted as genuine agents insofar as they satisfy the three conditions proposed as necessary and jointly sufficient for agency. What all agents from the simplest to the most sophisticated ones have in common is the intrinsic goal to be; in other words, the continuation of their very existence is what drives them first and foremost.

This self-organization approach avoids some of the difficulties evolutionary etiology approaches confront and can account, in particular, for the here and now goal-directedness of a system’s action. At the same time, it takes the idea that the overarching goal or norm of agency is self-maintenance as a primitive, as constitutive of what it is to be an agent. Whether we should accept that self-maintenance is axiomatic of agency or think this is an assumption in need of an explication remains a disputed question.

While my focus so far in this section was on goal-directedness without representation, this is not to say that animals are unable to represent goals. Here again we should accept the idea of a representational continuum. Along this continuum we find animals that can represent goals perceptually. Perception may set an object for animal action. A cat seeing a mouse will try to catch it. Memory can also be used to set goals. A thirsty animal who has an internal map of its territory stored in memory may use it to set an action goal, such as going to the closest water hole (Kennedy & Shapiro, Reference Kennedy and Shapiro2009). Of course, as Burge points out, at low levels of representational agency, an animal “does not set its goals as good, or as good for it. It does not reason about its goals. It lacks propositional intentions” (2009: 277). Such feats are only possible for creatures endowed with the kind of representational capacities that become available once we reach the high end of the representation continuum.

2.3 Agentive Control

Many philosophers also insist that it is a necessary condition of agency that the putative agent have control over their action. For instance, Frankfurt (Reference Frankfurt1978) criticizes the claim, defended in early versions of the causal approach, that the essential difference between what an agent does and mere happenings lies in their prior causal histories. Frankfurt forcefully argues for the view that guidance or control of bodily movements by the agent as the action is being performed is central to agency and that the conditions for attributing the guidance of bodily movements to a whole creature and not just some subsystem within the creature can obtain for many creatures besides humans, including, in his own example, spiders. We can distinguish between two dimensions of control. First, agents should have control over whether they act and what they do. Second, agents should also have control over how they do what they do, that is, be able to guide and shape action execution.Footnote ²

There are dissenters to the view that agential control is a necessary condition of agency. For instance, Burge (Reference Burge2009: 264) contends that control is not central to primitive agency. He claims that ducking an approaching missile can count as an action, even if the ducking is not something the individual can inhibit, at least provided that the ducking is not a mere peripheral reflex but a whole-organism behavior and has a biological function (e.g., avoiding harm). Similarly, he claims that primitive whole-organism agency “does not require a capacity to shape or guide whole-organism movement past the point where the stimulus is registered” and that “various types of instinctive behavior are inflexible and chain-reflexive, but still count as action” (2009: 265). These are borderline cases, and some authors (e.g., Ledoux & Daw, Reference LeDoux and Daw2018) prefer to classify reflexes and fixed action patterns as reactions rather than as actions.

The first dimension of control is not just about whether to act or not; it is also about how organisms “decide” which goal, from a possible set of goals, to pursue. This issue arises even for creatures with only primitive or minimal agency. Whether we conceive of goals in terms of biological functions as understood by Burge and the evolutionary etiology approach or in terms of self-maintenance as in self-organization theory, we still need to explain how an organism settles on one goal rather than another. In particular, the behavior of an organism does not just depend on what the environment has to offer; it also depends on its internal states, needs or drives. In addition, the value of a behavior does not just depend on the specific drive or need a stimulus satisfies; it is also modulated by other factors such as the other current needs or drives of the organism.

In the animal literature on motivated behavior, the classical theories of motivated behavior regulation are homeostatic theories. According to homeostatic theories, motivated behavior seeks to keep internal drives in equilibrium and restore balance among different physiological needs (such as warmth, food, and hydration). A central idea of homeostatic models is that homeostatic regulation is based on neurobiologically encoded setpoints corresponding to levels that are good for survival together with interoceptive signals that report current homeostatic levels. If we take the need for food as an example, the setpoint might correspond to a certain level of glucose in the blood and an interoceptive signal might report the current level of glucose in the blood. Whether the difference between current level and setpoint steers action (in this case, feeding) depends on other homeostatic levels (for instance, an organism may be seriously dehydrated and in more urgent need of water than of food). In other words, we may understand needs and drives as dimensions of a value space and motivated behavior as an attempt to minimize the maximum distance to setpoints in this value space.

It has been proposed that homeostatic models of physiological need regulation should be supplemented by reinforcement learning models to account for the richer set of goals and learned setpoints of more complex organisms capable of more flexible behavior (e.g., Keramati & Gutkin, Reference Keramati, Gutkin, Shawe-Taylor, Zemel, Bartlett, Pereira and Weinberger2011). Recently, Juechems and Summerfield (Reference Juechems and Summerfield2019) have proposed that the homeostatic approach could also be extended to account for human motivated behavior in cognitive, economic, or social settings by understanding this behavior as seeking to restore a balance not just among competing physiological needs but also among competing acquired cognitive setpoints that correspond to higher-order goals.

While it would be nice to have a unified theory of motivated behavior applying to all agents from the simplest organisms to humans, the extension of the homeostatic approach to human behavior raises a worry. It assumes the existence of a unique multidimensional value space the dimensions of which variously pertain to both physiological states and cognitive goals and of a process that operates on this multidimensional space and optimizes over states by minimizing distance to the setpoints in the space. Such a model can account for indifference between options if there are several equally good ways to minimize distance to the setpoints in the space. Yet, it cannot straightforwardly account for motivational and decision conflicts. For instance, there is a long philosophical tradition of investigating akrasia, classically defined as involving a conflict between our better judgments on what to do and motivational forces that drive or compel us to act in opposing ways (e.g., judging it best to study for one’s exams while being more motivated to watch your favorite show on television and, acratically, remaining in front of the television).

To account for the (very real) possibility of akratic actions, some philosophers have argued that we need to drive a wedge between evaluation and motivation, allowing that the action one evaluates as best need not be the action one is most motivated to perform. For instance, Mele (Reference Mele1987) argues that our judgments of what we should do are based at least in part on our evaluation of the “objects” of our desires rather than on the motivational force of the desires themselves. Thus, our evaluative judgments should not be seen as mere reflections of the overall balance of our motivational forces. The value we attribute to the objects of our desires may be out of kilter with the motivational force of these desires. It is the misalignment of evaluation and motivation that makes room for akratic action. This suggests that, at least as far as the explicit cognitive goals of humans are concerned, decision processes may rest on other computations (i.e., evaluative judgments) than just those proposed by homeostatic models.

The second dimension of control concerns action execution, how we guide, shape, monitor, and, if need be, correct or adjust its course. Even if we accept with Burge that some forms of primitive agency do not allow for action control in this sense, many organisms with low-level representational agency are capable of some form of control over action execution. They will adjust their movements in response to environmental disturbances (e.g., a gust of wind) or make a detour to reach their destination when encountering an unexpected obstacle.

In the computational literature on motor control, the so-called comparator model has played a very influential role. According to this model, motor control involves two main kinds of internal models, forward and inverse models, as illustrated in Figure 1. In a nutshell, an inverse model (or controller) computes the commands for achieving a desired state given the current state of the system and of the environment. An efference copy of these commands is fed to a forward model (also called a predictive model) that simulates how the system would behave over time if these commands were executed and can thus generate a prediction of the consequences of performing these commands. Of special interest is the idea that the control of action depends in a large part on the coupling of inverse and forward models through a series of comparators, that is, mechanisms that compare two signals and use the result of the comparison for various kinds of regulation. A first kind of comparator (labeled A in Figure 1) takes as input representations of the desired state and of the predicted state and sends an error signal to the inverse model if a difference is found. Such a mechanism can be used to maintain accurate performance in the presence of delays in sensory feedback. It can also be used for mental practice and planning, as forward models can predict the sensory outcome of an action without the action being actually carried out. A second kind of comparator mechanism (labeled B in Figure 1) compares the predicted consequences of a motor command with its actual consequences. The result of this comparison can be used to update the forward model and improve its functioning. It can also be used to filter sensory information and to distinguish the components that are due to self-movement from those due to changes in the world (Blakemore et al., Reference Blakemore, Wolpert and Frith2000). Finally, a third kind of comparison is between desired state and actual feedback (labeled C in Figure 1). Errors derived from the difference between the desired state and the actual state can be used to update the inverse models and improve performance. This kind of comparison is therefore important for motor learning.

Figure 1

The basic components of a motor control system based on internals models.

While the internal model theory of motor control was initially introduced to account for fine-grained aspects of motor control (such as joint angles, torque, limb positions, and trajectories), more recent versions of the theory emphasize the hierarchical nature of motor (or better action) control (Jeannerod, Reference Jeannerod1997; Grafton & Hamilton, Reference Grafton and Hamilton2007; Botvinik, Reference Botvinick2008). They propose that internal inverse and forward models are arranged in a hierarchy and that error signals generated at one level of the hierarchy can propagate to the next level when correction mechanisms at the former level cannot make the necessary compensations.Footnote ³ As a first approximation, we may distinguish three main levels in this control hierarchy corresponding to the three levels of the DPM model of intentions discussed in Section 1.5, with inverse and forward models operating at different levels of grain associated with each level.

Acknowledging the existence of different levels of action control may allow us to accommodate both automatic and non-automatic action control processes. Sensorimotor control at the level of M-intentions operates automatically and is responsible for fast, non-conscious adjustments of movements. In contrast, control processes at higher levels corresponding to P-intentions and D-intentions operate at a longer time-scale and are typically conscious. Indeed, direct evidence for this duality of action control processes in humans comes from a set of experiments (Fourneret & Jeannerod, Reference Fourneret and Jeannerod1998; Fourneret et al., Reference Fourneret, Franck, Slachevsky and Jeannerod2001, Reference Fourneret, Vignemont and Franck2002) in which participants were instructed to move a stylus on a graphic tablet along a straight line to a visual target. Participants could not see their drawing hand, but its trajectory was visible as a line on a computer screen. On some trials, the experimenter introduced a directional bias electronically so that the visible trajectory no longer corresponded to the trajectory of the hand. When the bias was small, participants made automatic adjustments of their hand movements to reach the target but remained unaware that they were making these corrections. It is only with larger biases that participants became aware of a discrepancy and began to use conscious correction strategies to compensate for the bias and reach the target.Footnote ⁴ These results suggest that although small discrepancies between predicted and actual sensory feedback are detected at some level since they are used to make appropriate corrections of the hand movement, they are not normally consciously monitored and the corrections are automatic, involving only the intervention of sensorimotor control. However, when discrepancies become too large to be automatically compensated, we become aware of them and shift to a conscious compensation strategy corresponding to situational control. In addition, human agents can also exert rational control over their actions to ensure that the way the action unfolds – in particular, unanticipated side effects or changes to the action plan in the face of unexpected difficulties – doesn’t flout the reasons the agent had for their action in the first place or violate norms of consistency and practical rationality. For instance, if a few minutes after I started jogging intending to do the full circuit around the lake, I suddenly remember that I have an important videoconference in less than half an hour, rational control dictates that I change my plan and do a shorter circuit allowing me to be back in time for the videoconference.

The examples of sensorimotor, situational, and rational action control I just gave all involved human actions. Rational control requires high-level representational agency, with propositional attitudes, language-like representations and the inferential promiscuity they allow and, thus, may exceed the agentive capacities of most animal species. Yet, many animal species may be capable of situational control and most of them may have some form of sensorimotor control over their action. In the same way that we should think of representational abilities as forming a continuum rather than in terms of simple representational/non-representational dichotomy, we should think of action control as forming a continuum. At the low end of the continuum, we find organisms with very primitive agency and no capacity for control. At the high end, we find organisms capable of the three forms of action control, sensorimotor, situational and rational I just described. In between these two ends, we find organisms with capacities for more limited forms of action control.

So far, our discussion of the minimal conditions for agency has focused on the agency of living organisms. Taking stock, we have considered three main requirements on agency. First, agency requires sensitivity to environmental conditions, where this sensitivity falls along a continuum from non-representational sensing of stimulus conditions (non-mental sensings, in Burge’s terminology) all the way to full-fledged beliefs about the world. While many would take it that non-human animals lack the representational and computational resources needed for the ascription of propositional attitudes and for inferential promiscuity, many animal species may still be capable of representing their environment using a variety of less sophisticated representational formats. Second, agency requires purposiveness or goal-directedness: To count as an action of an organism behavior must be endogenously driven and its purpose or goal be intrinsic and attributable to the whole organism and not just to some subsystem within it. Again, we have seen that goal-directedness forms a continuum from goals that are not represented but correspond to some biological function of the organism all the way to propositional intentions. Thus, even if they don’t have propositional intentions, many animal species may form representations of goals using representational formats that are less sophisticated than those needed for propositional intentions. Third, agency may require that agents have some control over their actions, over whether they act or not, what they do and how they do it. While Burge contends that this condition is not central to primitive forms of agency, others would take action control to be a condition on genuine agency. Once again, we have seen that action control can take more or less sophisticated forms in the animal realm.

2.4 What About Artificial Agents?

Let me conclude with some brief remarks on artifacts. Can some artifacts qualify as genuine (artificial) agents? First, in the same way that biological agency is not a homogeneous category, artificial agency is not a homogeneous category, a mere thermostat regulating the temperature in a room, if it is an agent at all, is worlds apart from the latest and most sophisticated generation of artificial intelligence agents (AI agents for short). But let us focus on AI agents as the best potential candidates for genuine agency.

In their influential and comprehensive overview of AI research, Russell and Norvig (Reference Russell and Norvig2009) define AI as “the study of agents that receive percepts from the environment and perform actions” (Russell & Norvig, Reference Russell and Norvig2009: vii). From its inception in the 1950s, work in the areas of planning and problem solving has been central to research in AI. For instance, an important approach to planning in AI, the BDI architecture,Footnote ⁵ was inspired in part by Bratman’s theory of planning agency (e.g., Rao & Georgeff, Reference Rao and Georgeff1995). Early AI planning and reasoning systems were tailor-made to narrow and well-circumscribed domains. However, the exponential growth of AI in the last two decades, the wealth of new formalisms and new computational and learning techniques has allowed AI to overcome most of these limitations. In terms of representational and reasoning capacities, some of the current AI systems are now at the high-end of the representation continuum and clearly meet the first condition on agency (for recent surveys of advances in the reasoning and representational abilities of AI systems, see, e.g., Bhuyan et al., Reference Bhuyan, Ramdane-Cherif, Tomar and Singh2024; Sun et al., Reference Sun, Zheng and Xie2025).

With respect to action control, the comparator model we discussed in Section 2.3 was inspired by the work of engineers who proposed computational theories incorporating the idea of control strategies based on internal models, applying this approach in the fields of robotics, neural networks, and adaptive control. Since then, the more recent predictive coding approach to brain mechanisms of action control has also gained a hold in robotics (e.g., Taniguchi et al., Reference Taniguchi, Murata and Suzuki2023). Thus, reasonably sophisticated AI agents appear to also meet the control condition on agency.

What is more problematic is the second condition of intrinsic, endogenous purposiveness or goal-directedness. While AI systems can certainly represent goals and reason about goals, the goal-directedness of the system appears to be extrinsic, imposed from the outside by the system’s designers. AI agents do not have biological goals resulting from the evolutionary history of the ancestors of the system; nor can an IA agent be considered a self-organizing system with self-maintenance and the continuation of its own existence as its ultimate goal.

Whether this is an insuperable obstacle to ascribing genuine agency to AI agents is at present unclear. For instance, evolutionary robotics is a branch of robotics using evolutionary principles (selection, variation, heredity) for the design of robots (Doncieux et al., Reference Doncieux, Bredeche, Mouret and Eiben2015). A pivotal feature of this approach is its holistic character: It considers the robot, its environment, and all the interactions between its components all at once. Whether after many generations, robots so designed will have acquired functions that qualify them as intrinsically goal-directed remains an open question. Another emerging paradigm in AI, Agentic AI, works on the development of autonomous systems designed to demonstrate adaptability, advanced decision-making capabilities and self-sufficiency and to pursue complex goals with minimal human intervention (Acharya et al., Reference Acharya, Kuppan and Divya2025). At present, it might be best to consider existing AI agents as borderline cases, meeting some but not all the conditions of genuine agency. Yet, with the explosive growth of AI in general and large language models in particular, this position may have to be reevaluated in a not-so-distant future.

3.1 Agentive Knowledge

It is often held that awareness from the inside provides a distinctive form of knowledge of at least certain aspects of our own agency. But what form of knowledge is it and which aspects of our agency does it concern? Traditionally, knowledge from within has been equated with introspective knowledge, the only possible objects of which are inner mental states or events. But then, knowledge from within could only be knowledge of our intentions and volitions, not of our bodily actions per se. Some philosophers have wanted to go further and claim that we have knowledge from within of our bodily actions themselves and not just their mental antecedents. Thus, in her book Intention (1957), Anscombe argues that intentions provide us with a special kind of self-knowledge and claims that this knowledge is special in two ways. It is knowledge of our own intentional actions, that is, knowledge not just of what one is attempting to do, but of what one is actually doing, and it is knowledge without observation. With these claims, Anscombe is trying to steer a middle course between two extreme stances she rejects. On the one hand, in insisting that our knowledge of our own actions is non-observational, she wants to draw a contrast between the knowledge we have of the actions of others, which is based on observation, and the knowledge we have of our own actions, which is a sui generis kind of self-knowledge. On the other hand, she also rejects as ridiculous the idea that agentive knowledge is a form of introspective knowledge. According to her, we know of our current actions not on the basis of observation or on the basis of introspection, but because we have non-observational knowledge of our practical reasons and our actions are the upshot of our practical reasoning. Put differently, knowledge of our intentions and reasons for acting is also knowledge of what we are doing, in virtue of the rational control intentions and reasons exert on our actions.Footnote ⁶ She argues that rather than as a form of speculative knowledge, agential knowledge should be understood as a form of practical knowledge. While speculative knowledge aims at describing the world, practical knowledge aims to bring something about the world. For instance, if I am intentionally painting the wall yellow, I know that I am doing so, not by observation, but in virtue of the fact that my intending to do so is the cause of my doing so. This knowledge is thus a form of practical knowledge.

There has been much philosophical debate on how exactly Anscombe’s claims should be interpreted and whether they should be accepted or rejected. Many philosophers have objected to her claim that that to act intentionally an agent must know what they are doing as too strong and seen her characterization of agential knowledge as practical knowledge as, at best, mysterious. In recent years, several authors have proposed weakenings of the knowledge condition on intentional action as well as deflationary interpretations of Anscombe’s claims about agential knowledge.Footnote ⁷ For instance, Falvey (Reference Falvey2000) and Velleman (Reference Velleman, Leist and Baumann2007) favor a reliabilist interpretation of Anscombe’s claims. According to this interpretation, knowledge of one’s own intentional actions is non-observational because it is given by our awareness of the content of our intentions and because intentions normally constitute (practical) knowledge of our own intentional actions insofar as they reliably cause the facts that make them true. Note that on this reliabilist reading, Anscombe’s claim is not that the content of our intentions provides us with infallible knowledge of what we are doing. To say that there normally exists a reliable connection between our intentions and actions is not to say that there cannot be cases when this connection does not obtain. However, as Velleman emphasizes, on a reliabilist interpretation of Anscombe’s account, failures of reliability undermine not just the epistemic status of intentions, they also undermine the intentionality of actions. If my intending to A does not reliably cause my A-ing, my intending to A will not amount to knowledge that I am A-ing, since my intention may not have caused me to A. Conversely, there is no guarantee that my A-ing when it happens is actually caused by my intention and is thus an intentional action. It may instead be an accident that I intend to A and that I also happen A.

While they grant that there are important connections between action control and knowledge of action, many philosophers reject Anscombe’s account of this connection. Instead, they consider our knowledge from within of our actions to be grounded in our experience of acting. A proper delimitation of this knowledge would thus require a clear determination of the contents of this experience and of its relation to action control (e.g., Peacocke, Reference Peacocke, Roessler and Eilan2003). We now turn to these issues.

3.2 Agentive Awareness

Agentive awareness, often called sense of agency, comprises both agentive beliefs and agentive experiences (Bayne & Pacherie, Reference Bayne and Pacherie2007).Footnote ⁸ Agentive experience – that is, the specific subjective experience one might have when one is performing or about to perform an action – constitutes our basic form of agentive awareness.Footnote ⁹ Our agentive beliefs, while often based on our agentive experiences, can also be based on more indirect evidence. For instance, if I find my car keys in the fridge, I may form the belief that I was the one who put them there, despite having no recollection of having done so. In this case my agentive belief will be based on other doxastic states, such as my belief that I was alone at home and perhaps my further belief that I have a pronounced tendency toward absent-mindedness, rather than on any specific agentive experience or memory thereof.

For now, my primary focus will be on agentive experiences, although I will return to agentive beliefs later in this section. The burgeoning literature on the content and sources of agentive experiences highlights their many facets. A non-exhaustive list of aspects or components of agentive awareness would include: awareness of a goal, awareness of an intention to act, awareness of initiation of action, awareness of movements, awareness of activity, experience of effort, experience of control, experience of authorship, experience of intentionality, experience of purposiveness, experience of trying, experience of freedom, and experience of mental causation. Mylopoulos and Shepherd (Reference Mylopoulos, Shepherd and Kriegel2020) have recently proposed a taxonomy that distinguishes six clusters of properties or aspects of agentive phenomenology: the phenomenology of purposiveness, the phenomenology of mineness, the phenomenology of execution, the phenomenology of action perception, the phenomenology of action assessment, and the phenomenology of free will.

Another way to avoid becoming overwhelmed by the profusion of aspects of agentive experiences is to distinguish, following Bayne (Reference Bayne, Aguilar, Buckareff and Frankish2010), core and non-core elements of agentive experiences. Core elements are elements that must be possessed by any agentive experiences whatsoever and that are essential to agentive experiences, and non-core elements are elements that may but need not be present within experiences of agency.

Among non-core elements of agentive experiences, one may include, among other things, experience of effort, experience of deliberation, or experience of decision. Arguably, not all actions are effortful or felt as such. Indeed, one finds reports in the clinical literature of people who after brain damage have lost the capacity to experience actions as effortful but still enjoy agentive experiences (Naccache et al., Reference Naccache, Dehaene and Cohen2005). It is also not uncommon for one to act on the spur of the moment, or for actions, especially routine actions, to be triggered more or less automatically by cues in the environment. We can have agentive experiences for these actions, but these won’t include experiences of deliberation or of decision. This is often the case for routine actions that can be triggered more or less automatically. For instance, I found myself a few minutes ago cleaning my glasses. As soon as I became aware of what I was doing, I had an agentive experience for the action I was performing but it did not include among its elements an experience of deliberating whether or not to clean my glasses or even deciding to do so.

Other elements, however, are essential to agentive experiences and in what follows I will concentrate on these core elements. At the very minimum, to have an agentive experience is to be aware of oneself as acting, that is to experience a sense of agency, narrowly conceived. Note that in saying that the sense of agency narrowly conceived is a core component of agentive experiences, I am not implying that this core component is itself necessarily atomic. Indeed, it is often proposed that the experience of agency is itself decomposable into at least an experience of initiation and an experience of control.

It is difficult to conceive of an experience of agency as existing independently of any experience of what it is one is doing, as a feeling simply floating in the air, so to speak. It seems that an agentive experience also minimally involves some sense of what it is one is doing, whether it be an experience of moving one’s body in a certain way (e.g., moving a finger), of performing an action of particular kind (e.g., pressing a key) or of acting with a certain goal (e.g., starting a new paragraph). Having an agentive experience does not require one to identify the action the experience is about at a particular level of specification, but it demands that the action be specified in some way. Thus, I take it that agentive experiences include two core elements: a sense of oneself as acting and some awareness, however thin, of the action one is performing.

Note that this preliminary regimentation leaves open a number of questions: How are these various components of the phenomenology of agency related? Are agentive experiences veridical? Do they play a causal role in shaping one’s actions and, if so, what role? To answer these questions, we need to consider what the sources of agentive experiences are and how the mechanisms responsible for generating them relate to the mechanisms involved in action production and control. To this we now turn.

3.3 Sources of Agentive Experiences

Empirical research on the sense of agency has been very active in the last three decades and has uncovered a rich set of mechanisms and factors contributing to the emergence and modulation of agentive experience. A number of different models of how agentive experiences are generated have been proposed. Typically, these models focus more on core elements of agentive experiences than on non-core elements. Early research on the processes underlying the sense of agency was framed by the opposition of two theoretical positions: the motor prediction view and the cognitive reconstruction view. More recent research has shown that the two views were not mutually exclusive and that both motor prediction and cognitive reconstruction contribute to the sense of agency (Moore & Haggard, Reference Moore and Haggard2008).

The motor prediction view, or comparator model approach, is inspired by computational theories of motor control, discussed in Section 2.3. In the 1990s and early 2000s, Chris Frith and his colleagues developed their very influential comparator model of the sense of agency (Frith et al., Reference Frith, Blakemore and Wolpert2000). They proposed that the signals used for motor control also provide cues to agency: Whether one has a sense of agency for an action and how strong this sense of agency is depends on the degree of congruence between desired state, predicted state, and estimated actual state (Frith et al., Reference Frith, Blakemore and Wolpert2000). In particular, this model proposed that awareness of initiating an action is based on a representation of the predicted consequences of making that action, rather than its actual consequences, and on the congruence between the predicted state and the desired state. It also proposed that for this experience of agency to persist, the predicted consequences would also have to remain congruent with the sensory reafferences when they become available. These two claims are associated with the idea that the sense of agency has two main components: the sense of initiation linked to the congruence between predicted and desired state and the sense of control linked to the degree of congruence between predicted and actual state. In addition, the predicted sensory consequences of an action are typically attenuated, feeling less intense than the same sensory events produced by external sources. This allows agents to distinguish between the sensory consequences of their actions and sensory events that have independent causes. This view is empirically supported by multiple studies showing, on the hand, that spatial and temporal discrepancies between the expected and perceived consequences of an action reduce the sense of agency for that action (e.g., Knoblich & Kircher, Reference Knoblich and Kircher2004; Sato & Yasuda, Reference Sato and Yasuda2005) and, on the other hand, that the sensory consequences of voluntary actions are attenuated in healthy participants (e.g., Blakemore et al., Reference Blakemore, Wolpert and Frith2000; Gentsch & Schütz-Bosbach, Reference Gentsch and Schütz-Bosbach2011).Footnote ¹⁰

In contrast, the cognitive reconstruction view downplays the contribution of the motor system to the sense of agency and proposes that it is the result of interpretative processes, distinct and separate from the processes involved in the specification and control of the action. Thus, on Wegner’s “theory of apparent mental causation” (Wegner, Reference Mele and Nadel2002), a general-purpose causal inference process is at play. If an observed action outcome is consistent with a prior thought of the agent and other potential causes of the action are not present or salient, a sense of agency for the action will be induced.

There is also empirical evidence that high-level inferential processes play a role in determining the sense of agency for an action. Studies by Wegner and colleagues have shown that cognitive cues can alter the sense of agency for an action independently of changes in sensorimotor and perceptual cues. For instance, in their “I-Spy” study (Wegner & Wheatley, Reference Wegner and Wheatley1999), a participant and a confederate of the experimenter had joint control of a computer mouse that could be moved over any one of a number of pictures on a screen. When participants had been primed with the name of an item on which the mouse landed, they expressed a stronger sense of agency for the action of stopping on that object (when in fact the stop had been forced by the confederate). In another experiment, the “helping hand” study, Wegner, Sparrow, and Winerman (Reference Wegner, Sparrow and Winerman2004) also demonstrated that it was not even necessary for a subject to actually move to experience a sense of agency. They had participants watch themselves in a mirror while another person behind them, hidden from view, extended hands forward on each side where participants’ hands would normally appear and performed a series of movements. When participants could hear instructions previewing the movements, their sense of agency for these movements was enhanced, but such vicarious agency was not felt when the instructions followed the movements. In addition, even when such previews are unavailable, people may engage in post-hoc confabulation. As Wegner puts it: “Even when we didn’t know what we were doing in advance, we may trust our theory that we consciously will our actions and so find ourselves forced to imagine or confabulate memories of ‘prior’ consistent thoughts” (2002: 146).Footnote ¹¹

There has been a growing consensus that the motor prediction view and the cognitive reconstruction view are not mutually exclusive but complementary. This reconciliation was helped in part by the introduction of a distinction within the sense of agency between agentive experiences and agentive judgments mentioned earlier. Agentive experiences can be experienced pre-reflectively, corresponding to a non-conceptual feeling of control and are thought to rely primarily on predictive processes and sensorimotor cues. Agentive judgments correspond to higher, conceptual level representations of oneself as the author of an action and rely on a larger set of cues and processes.

The currently dominant integrative framework appeals to hierarchical Bayesian models, where agency cues are weighted according to their reliability and prior beliefs contribute to and shape the emergent sense of agency (see Haggard & Chambon, Reference Haggard and Chambon2012; Moore & Fletcher, Reference Moore and Fletcher2012; Pacherie, Reference Pacherie, Haggard and Eitam2015). An important feature of the Bayesian hierarchical approach to the sense of agency is that, as we move up in the hierarchy, we also move away from predictive processes specifically dedicated to motor control toward more general-purpose predictive processes relying on a wider range of cues and priors. Relatedly, as we move up the hierarchy, we also move away from agentive experiences toward agentive judgments. In addition, recent work has also shown that beyond cues and processes associated with action programming and performance, the sense of agency is also influenced by internal processes involved in the selection of actions, in advance of the action itself. In particular, there is evidence that the fluency of action selection processes prospectively contributes to the sense of agency (Wenke et al., Reference Wenke, Fleming and Haggard2010; Chambon et al., Reference Chambon, Sidarus and Haggard2014; Sidarus & Haggard, Reference Sidarus and Haggard2016). Thus, when participants in an experiment have to select between two action options and the selection is facilitated by the presentation of a subliminal prime indicating the correct option, participants report a stronger sense of control over the effect of the action, suggesting that a more fluent action selection process increases their sense of agency. Finally, properties of action outcomes have also been shown to influence the sense of agency for an action. For instance, it has been shown that that the magnitude of an action outcome also strongly influences the sense of agency (Kawabe, Reference Kawabe2013) and that the emotional valence of the outcome also modulates the sense of agency (Tanaka & Kawabata, Reference Tanaka and Kawabata2021).

3.4 Veridicality of Agentive Awareness

How accurate are our agentive experiences and judgments? Are they faithful reflections of what we actually do? In principle, there are a number of ways in which things could go wrong. We might have a sense of agency for actions that were not actually ours or we might lack a sense of agency for actions that we actually performed. We might also be mistaken about what exactly we did and why. Are such mistakes the exception or rather the rule? Can our sense of agency be considered generally reliable even if it can occasionally go astray or is it a systematic illusion?

There are various pathologies or conditions where the sense of agency is altered. Delusions of control, one of the main symptoms in the constellation of positive symptoms diagnostic of schizophrenia, is characterized by alterations of the sense of agency. People with delusions of control experience some of their actions as controlled by an external agent (e.g., Spence et al., Reference Spence, Brooks, Hirsch, Liddle, Meehan and Grasby1997; Frith, Reference Frith2005) but can also self-attribute actions performed by others (e.g., Maeda et al., Reference Maeda, Kato and Muramatsu2012). Patients with anosognosia for hemiplegia following brain damage claim to have performed the action when asked to move their paralyzed limb and report a sense of agency for the inexistent movement (Piedimonte et al., Reference Piedimonte, Garbarini, Pia, Mezzanato and Berti2016). Under hypnosis, people readily respond to motor suggestions but report a compelling experience of passivity with regard to their actions (Martin & Pacherie, Reference Martin and Pacherie2019). Such cases are typically explained in terms of a more or less severe and perduring disruption of the processes and mechanisms responsible for the generation of the sense of agency.

Yet, what about healthy subjects where these processes and mechanisms are unimpaired? Are their agentive experiences and judgments reliable guides to their agency? How we answer this question depends on what we take their contents and sources to be.

According to the cognitive reconstruction view and to Wegner’s theory of apparent mental causation (Wegner, Reference Mele and Nadel2002), the sense of agency is the result of interpretative processes, distinct and separate from the processes involved in the specification and control of the action. Wegner thinks that his theory provides evidence that agentive experiences are illusions. In calling them illusions, he seems to mean both that our experiences are often non-veridical – they misrepresent the agent and the structure of their actions – and that they are epiphenomenal – they play no causal role in the production of the action they are experiences of. We will come back to the causal efficacy or lack thereof of agentive consciousness in the next section, but for now let us focus on the issue of veridicality. Wegner takes our agentive experiences to represent our conscious intentions as causing our actions. He sees the non-veridicality of our agentive experiences as a consequence of the fact that the processes through which the experience of agency is generated are quite separate from the processes involved in action-specification and control. Wegner seems to think that our experiences of acting could only be veridical if they had some direct link to the processes actually involved in the production of action; instead, their being, according to his model, theoretically mediated makes them illusory. As several commentators have noted, even if one were to grant Wegner’s claim that experiences of agency are generated by retrospective inferential processes, his view of what the contents of agentive experiences include is highly loaded and quite contentious and it is unclear why inferential mediation should be incompatible with reliability (e.g., Bayne, Reference Bayne, Pockett, Banks and Gallagher2006; Gallagher, Reference Gallagher2007).

In contrast, the comparator model of the sense of agency takes the mechanisms responsible for the generation of agentive experiences to be closely related to the mechanisms involved in action production and control. On this model, provided the mechanisms in question are functioning normally, agentive experience should accurately reflect our actual agency. However, as we have seen, both the cognitive reconstruction model and the comparator model offer only partial models of the mechanisms underpinning the sense of agency. More recent integrative models, including Bayesian hierarchical models, suggest that both predictive and retrospective processes contribute to the sense of agency, that predictions may be more or less fine-grained and precise, and that the weight given to agency cues at different levels of the hierarchy of representations involved in shaping agency is a function of their reliability. This invites a more nuanced answer to the question of the veridicality of agentive experiences and judgments. As we saw in the previous section, as we move up in the hierarchy, we also move away from predictive processes specifically dedicated to motor control toward more general-purpose predictive processes relying on a wider range of cues and priors. Conversely, as we move down in the hierarchy, we move toward unconscious motor representations and automatic control processes we have no awareness of. The line drawing experiment presented in Section 2.3 illustrates the lower limits of agentive awareness and at the same time the limits to the accuracy of our experience of our own movements. As we move toward the upper end of the hierarchy, we may encounter other kinds of limitations. We are not always aware of the reasons why we act as we do and when asked the question “Why?” may readily engage in confabulation. One illustration of this phenomenon is the choice-blindness effect. For instance, in an experiment (Johansson et al., Reference Johansson, Hall, Sikstrom and Olsson2005) participants were presented with pairs of pictures of faces and asked to choose between them on the basis of attractiveness. Once they had made a choice, they were presented again with the picture they had chosen and asked to provide reasons for their choice. Unbeknownst to them, on certain trials the experimenters used a card magic trick to covertly exchange one face for the other. Not only did the majority of participants fail to notice the substitution, they also readily came up with reasons for the choice they had not actually made! Yet, between these lower and upper bounds, agentive awareness may show greater reliability.

3.5 Causal Efficacy of Conscious Agentive States

Do conscious agentive states play a causal role in shaping our actions or are they merely epiphenomenal? Skeptical arguments regarding the causal efficacy of conscious agentive states have had two main targets: the role of conscious intentions in the initiation of action and the role of conscious visual states in the control of an unfolding action.

3.6 Concluding Remarks: Widening Our Horizons

We have seen that there are many aspects of conscious agency, from our awareness of our intentions and reasons to act all the way to our immediate awareness of guiding and controlling an ongoing action. We have also seen that blanket skepticism about the veridicality of our conscious experiences of agency has little warrant and that agentive consciousness, while not infallible, is in general a reliable guide to our agency. Finally, it should be noted that skeptical arguments regarding the causal efficacy of conscious agency typically have a rather narrow focus.

Arguments for will-skepticism focus on the role of proximal intentions in-action initiation and fail to consider the potential role of conscious agency prior to action initiation. Yet, even if it were true that the subjects in Libet’s experiment initiated their movements when they did as a result of unconscious decisions, it is quite implausible that the participants in his studies would have produced these movements unless they had intended to participate in the experiment and to comply with the experimenter’s instructions in the first place. This leaves plenty of room for conscious agency to play a role in advance of action initiation.

Similarly, zombie action arguments narrowly target the role of conscious visual experience in the motor control of ongoing action. Yet, in assuming that the only form of visual control of an ongoing action is automatic control by unconscious visual representations computed in the dorsal pathways, they fail to acknowledge that multiple levels of action control are simultaneously at play during action execution. In particular, in the spirit of Frankfurt (Reference Frankfurt1978), it is important to emphasize that conscious control does not necessarily mean causal intervention: Action may be consciously controlled even though unconscious rather than conscious control mechanisms intervene to adjust movements, provided conscious control mechanisms would have kicked in, had automatic control proved insufficient. In addition, we shouldn’t see automaticity as an unbearable affront to our vanity as conscious agents. Rather than picturing automaticity and conscious agency as fiends vying for the same territory, it may be more fruitful to see them as complementary forces, even if their cooperation isn’t always easy and straightforward. If anything, what makes human agency unique is the richness and the complexity of their interactions and the flexible ways in which control can being passed up to conscious processes or down to automatic processes as the need arises.

Finally, the occurrent sense of agency we experience for actions we are currently preparing or performing may also contribute to our long-term sense of agency, that is, to our general sense of ourselves as agents, our sense of what we are capable of over time, and ultimately to a form of self-narrative where one’s past actions and projected future actions are given coherence by being integrated into a set of overarching goals, motivations, projects, and general lines of conduct.

4.1 Shared Intentions

Philosophical accounts of shared intentions are attempts to elucidate what it takes for agents to act in a jointly intentional manner. These accounts all agree that joint actions are more than mere summations of individual actions and that the intentionality in joint action cannot reduce to the intentionality of the individual actions that together contribute to the joint action. They all agree, therefore, that something more is needed on the part of individual agents for them to be said to be acting jointly. But what more is needed? This is where disagreements start.

Let me briefly consider the differences between the accounts of shared intentions respectively advocated by Bratman (Reference Bratman2013), Searle (Reference Searle, P. Cohen and Pollack1990), and Gilbert (Reference Gilbert2009, Reference Cloutman2013). For Bratman, the main discontinuity in agency is not between individual and shared agency, but rather between simple purposive agency and planning agency, with the new practical elements essential to it, namely intentions. Bratman develops a constructivist approach to shared intentions that exploits the conceptual and normative resources of his planning theory of individual agency. His basic thesis is that one can capture the interconnections among agents that are characteristic of shared agency by construing shared intentions as complexes of interlocking and interdependent intentions and other attitudes of individual agents. While the interconnected intentions Bratman’s basic thesis appeals to have special content that makes reference to the intentions of other agents, they are, qua mental attitudes, ordinary personal intentions. This is an important difference with Searle’s account according to which what is essential to shared agency is a special individual attitude of “we-intention” (Searle, Reference Searle, P. Cohen and Pollack1990). A “we-intention,” as Searle conceives of it, is a primitive kind of psychological attitude. Thus, for Searle, in order to account for shared agency, we need to introduce a basic new element in the furniture of the individual mind. In contrast, Bratman’s basic thesis is metaphysically parsimonious in appealing only to psychological kinds that are already needed for individual planning agency.

Bratman also argues that the social normativity characteristic of shared agency derives from the normativity already associated with individual planning agency. As we have seen in Section 1.5, according to his theory of intentions, individual intentions are subject to central norms of consistency, agglomeration, coherence, and stability. The intentions of individual participants, when they are interconnected in the way specified by the basic thesis, will normally, in responding to these norms of individual practical rationality, lead to the emergence of corresponding norms of social consistency, social agglomeration, social coherence, and social stability.

This contrasts with Gilbert’s plural subject theory of shared intention and her view that the fundamental phenomenon in shared agency is joint commitment with the obligations and entitlements inherent in it (Gilbert, Reference Gilbert2009, Reference Gilbert2013). For Gilbert, two or more agents become the plural agent of a shared intention by jointly committing to pursue a certain goal as a body. A web of obligations and entitlements forms the normative structure of plural subjects and the way the plural subject, once constituted, manifests its unity is by emulating as far as possible a single body that intends to pursue that goal. For Gilbert the idea of a joint commitment is a primitive social notion that does not admit of further reductive analysis. Similarly, the obligations and entitlements a joint commitment grounds cannot be understood as moral obligations and entitlements or as emerging from the norms associated with individual planning agency. Rather, they engage a sui generis kind of social normativity. There is, thus, on Gilbert’s account, a deep conceptual, metaphysical and normative discontinuity between individual and social agency.

Despite their divergences, these philosophical accounts are all cognitively highly demanding. They require that agents participating in joint actions have robust theory of mind skills and high-level reasoning abilities, since, in order to form shared intentions, they must be able to represent their partners’ intentions and other attitudes relevant to the joint activity and to reason about these mental states in order to adjust their own intentions and actions to those of their partners. In recent years, these accounts have come under criticism as both ill-suited to account for the joint action abilities exhibited by young children and non-human animals and incomplete insofar as they say very little about the processes involved in the online coordination of actions.

4.2 Minimalist Accounts of Shared Agency

Dissatisfaction with philosophical accounts of joint action in terms of shared intention has prompted the emergence of so-called minimalist programs. Minimalist accounts of joint action can take more or less radical forms, either retaining the notion of shared intention as central to joint action but trying to offer less demanding accounts of what counts as a shared intention (Tollefsen, Reference Tollefsen2005; Blomberg, Reference Blomberg2011; Pacherie, Reference Pacherie2011, Reference Pacherie2013) or to replace the notion of shared intention with that of a shared or collective goal (Butterfill, Reference Butterfill2012, Reference Butterfill and Kiverstein2017). Minimalists generally agree that joint actions can take different forms – some more sophisticated and involving robust shared intentions, and others more basic and less cognitively demanding.

While minimalists agree that acting jointly with other agents requires one to be able to track efficiently their actions and to plan and execute one’s own actions toward a common goal in light of what others are doing, they contend that, for some joint actions at least, it is possible to do so by relying on more modest representational and cognitive abilities than those assumed by classical philosophical accounts. For instance, Butterfill (Reference Butterfill2012) proposes that some joint actions involve sharing goals rather than sharing intentions, where sharing goals requires only an understanding of goal-directed actions and their common effects and where the function of shared goals is to coordinate goal-directed activities. Importantly, Butterfill uses the term “goal” to refer to an outcome, actual or possible, rather than to a mental state of an agent and contends that there are ways of characterizing actions as goal-directed that do not require attributing intentions or any other propositional attitudes to agents. For instance, one may do so by taking a teleological rather than a mentalistic stance, where taking a teleological stance involves tracking the relation between an action, its (future) outcome, and the current situational constraints (e.g., Gergely and Csibra, Reference Gergely and Csibra2003; Butterfill & Apperly, Reference Butterfill and Apperly2013). For a goal, understood in this sense, to be shared, the following conditions should be met: (a) There is a single goal toward which the actions of two or more agents are directed; (b) each agent expects the other agents to perform an action directed to the goal; and (c) each agent expects this goal to occur as a common effect of all of their actions directed to the goal, her own and the others.Footnote ¹⁴ Butterfill also makes it clear that his account of joint actions in terms of shared goals is not meant to replace accounts that appeal to shared intentions. Shared intentions would still be needed for complex joint actions that would require advance planning. But in cases where agents need to coordinate their activities rather than their plans, appealing to shared intentions rather than simply to shared goals may be superfluous.

4.3 Real-Time Action Coordination

The success of joint action requires not just that we coordinate our intentions but also that we coordinate our actions while acting. What are the processes supporting online, real-time coordination? Cognitive psychology studies of joint action typically focus on the perceptual, cognitive, and motor processes that enable individuals to flexibly coordinate their actions with others in real time. This psychological literature has grown too vast to allow for an easy summary. Here, I present only a sample of it.

Knoblich, Sebanz, and colleagues (Knoblich et al., Reference Knoblich, Butterfill and Sebanz2011; Sebanz & Knoblich, Reference Sebanz and Knoblich2021) distinguish between two broad categories of coordination processes, emergent, and goal-directed. In emergent coordination, coordinated behavior occurs due to perception-action couplings that make multiple individuals act in similar ways. One source of emergent coordination involves interpersonal entrainment mechanisms, that is processes in which one system’s rhythm or timing becomes synchronized with another system’s rhythm or timing. For instance, people sitting in adjacent rocking chairs will tend to synchronize their rocking behavior, even if the chairs have a different natural rocking tempo (Richardson et al., Reference Richardson, Marsh, Isenhower, Goodman and Schmidt2007). The perception of common or joint affordances can also lead to emergent coordination. A joint affordance is a case where an object affords action to two people jointly that it may not afford to each of them individually. Thus, a seesaw may afford action to two kids, but not to a single child. A third source of emergent coordination is perception-action matching, whereby observed actions are matched onto the observer’s own action repertoire and can induce the same action tendencies in agents who observe one another’s actions (Jeannerod, Reference Jeannerod1999; Rizzolatti & Sinigaglia, Reference Rizzolatti and Sinigaglia2010). It is likely that such processes make partners in a joint action more easily predictable, and thereby facilitate coordination in action. Importantly, however, emergent forms of coordination can operate independently of any shared goals, which may be altogether absent, and do not ensure by themselves that the agents’ actions track a shared goal.

In goal-directed coordination, agents plan their own motor actions in relation to the joint goal and also to some extent to their partners’ actions. As emphasized by Knoblich and colleagues (Reference Knoblich, Butterfill and Sebanz2011), shared task representations – that is, representations of the joint goal as well as of each partner’s expected contribution to that goal – play an important role in goal-directed coordination. Shared task representations provide control structures that allow agents to monitor and predict what their partners are doing, thus enabling interpersonal coordination in real time. Empirical evidence shows that having shared task representations influences perceptual information processing, action monitoring, control, and prediction during the ensuing interaction (Sebanz et al., Reference Sebanz, Knoblich, Prinz and Wascher2006). Thus, for instance, people tend to predict the sensory consequences not only of their own but also of other participants’ actions (Wilson and Knoblich, Reference Wilson and Knoblich2005) and to automatically monitor their own and others’ errors (van Schie et al., Reference van Schie, Mars, Coles and Bekkering2004). An important complement to the co-representation of tasks and actions is the co-representation of perception. Thus, joint attention provides a basic mechanism for sharing representations of objects and events and thus for creating a perceptual common ground in joint action (Tomasello & Carpenter, Reference Tomasello and Carpenter2007; Tollefsen, Reference Tollefsen2005). Joint attention can also allow agents to perform joint actions more efficiently. For instance, a study by Brennan and colleagues (Brennan et al., Reference Brennan, Chen, Dickinson, Neider and Zelinsky2007) demonstrated that co-agents in a joint visual search space were able to distribute a common space between them by directing their attention depending on where the other was looking and that their joint search performance was thus much more efficient than their performance in an individual version of the task. Finally, real-time coordination of actions may involve communication among agents. Yet, this communication is not necessarily verbal, as verbal communication may not be feasible or be too slow to inform one’s partner in time. Instead, an agent may modulate or exaggerate certain parameters of an instrumental action during execution to provide information to a co-agent (Sebanz & Knoblich, Reference Sebanz and Knoblich2021). For instance, the first violinist in a string quartet may exaggerate the upward movement of their bow to signal an attack.

What does all this tell us about the capacities needed for shared agency? First, for complex joint actions that require advance planning, shared intentions and the robust theory of mind skills and reasoning capacities necessary to their formation would be needed. Second, all joint actions, whether planned in advance or not, also rely on a variety of other coordination processes to ensure action coordination during execution.

4.4 Sense of Shared Agency

Relative to individual action, the sense of agency for joint action has received much less attention. It has been suggested that when people coordinate their actions with others they may experience a sense of shared agency, that is, a sense that “we” did it (Pacherie, Reference Pacherie and Seemann2012).

In Pacherie (Reference Pacherie2014), I proposed that while the mechanisms supporting the sense of agency in individual action may also be involved in the emergence of a sense of shared agency, the specificities of joint action introduce further complexity. To begin with, joint action does not only require predicting the consequences of one’s own actions (self-predictions) but also predicting the actions of one’s partner and their consequences (other-predictions) and integrating both types of predictions to build predictions about the joint consequences of the combination of actions (joint-predictions). However, it is unclear whether low-level sensorimotor predictions and discrepancies between predictions and sensory feedback play as important a role in the sense of shared agency as they do for the sense of individual agency. Some recent studies (Dewey et al., Reference Dewey, Pacherie and Knoblich2014; Le Bars et al., Reference Le Bars, Devaux, Nevidal, Chambon and Pacherie2020, Reference Le Bars, Bourgeois-Gironde and Wyart2022) support the idea that agency cues are weighted differently for individual and for shared agency, with the sense of agency for individual action relying more on cues linked to sensorimotor predictions and the sense of shared agency giving more weight to cues derived from higher-level predictions.

In addition, joint actions can take a great diversity of forms and present various forms of asymmetries among agents. These asymmetries may be found at the action selection stage (e.g., decider vs. follower); during action execution, with possible qualitative and quantitative differences in the agents’ respective contributions to joint action (e.g., experts vs. novices; more central vs. more peripheric roles) and in the forms of coordination involved (asymmetric vs. mutual), as well as in respect to possible differences in how much each participant benefits from the action outcome, once it is completed (e.g., equal, proportional, or unfair distribution of benefits). Much remains to be done in exploring the impact these asymmetries might have on the emergence of a sense of shared agency. Existing studies suggest however that asymmetries tend to have a detrimental effect on the sense of shared agency (van der Wel, Reference van der Wel2015; Bolt et al., Reference Bolt, Poncelet, Schultz and Loehr2016; Le Bars et al., Reference Le Bars, Devaux, Nevidal, Chambon and Pacherie2020, Reference Le Bars, Bourgeois-Gironde and Wyart2022). For instance, Bolt and colleagues (Reference Bolt, Poncelet, Schultz and Loehr2016) investigated the effect of mutual coordination on the sense of shared agency using a task where people had to coordinate to produce a sequence of tones. They found that their participants experienced a stronger sense of shared agency in a version of the task where they had to produce the tones in alternation, which required tight mutual coordination, compared to a sequential version of the task where only the second partner had to adapt the timing of their actions to the timing of their partner’s actions.Footnote ¹⁵

Finally, while there is evidence that the emotional valence of an action outcome modulates the sense of agency for individual actions, this and other emotional factors may also modulate the sense of agency for joint action. First, joint action may help us achieve positive outcomes that it would be difficult or even outright impossible to achieve alone. In addition, as emphasized by Salmela and Nagatsu (Reference Salmela and Nagatsu2017), acting jointly may also be intrinsically rewarding insofar as it contributes to the emergence of positively valenced shared emotions, including social emotions such as fellow-feeling or sense of affiliation. Consider, for instance, two people preparing a meal together. They might not do it not only, or not so much, because it makes achieving the desired outcome easier and faster (sometimes it is simpler to be alone in a kitchen!) as because they enjoy doing something together.