Intelligence Offloading and the Neurosimulation of Developmental Agents

Cognitive offloading occurs when environmental affordances expand cognitive capacity while facilitating spatial and social behaviors. Capacity-related constraints are also important, particularly as embodied agents come online during development. Vast differences in brain size and offloading capacity exist across the tree of life. We take from multiple perspectives to understand the proportional contributions of internal models (brain) and externalized processing (offloading) in developing embodied computational agents. As developing nervous systems scale with body size and/or functional importance, offloading is also driven by neural capacity. Cognitive capacity is ultimately determined by various innate and environmental constraints. We propose a similar model for computationally developing cognitive agents. A regulatory model of cognition is proposed as a means to build cognitive systems that interface with biologically-inspired substrates. Multiple tradeoffs result from energetic, innate, and informational constraints, and determine the proportion of internal to external information processing capacity. As growth of a biologically-inspired substrate accelerates or decelerates over developmental time, it changes the acquisitional capacity of the agent. Our agent’s capacity limitations determine externalization potential, which is characterized by three parameters and two mathematical functions. The neurosimulation approach to intelligence offloading can be applied to a broad range of agent-based models and Artificial Intelligences.


1.
Introduction The interactions between the nervous system and environment are an important component to simulating intelligent behavior.While environmental realism (in the form of enrichment) is certainly important [1,2], some aspects of cognition are embedded in the environment itself.Let us consider the case of phototaxis in single-celled organisms.Light is acquired from the external environment, and serves as information for the internal nervous system.All of the computation is performed in the cell: morphological asymmetries and properties of the cell body are of primary importance [3].In the algae Chlamydomonas reinhardtii, the cell is steered towards or away from a light source by its flagellum based on the accumulation of light signals [4].In algae, this integration occurs via biochemical processes.Among marine annelids such as polychaetes [5] and the genus Platynersis [6], biochemical control of simple behaviors is replaced by small connectomes of 68 and 71 neurons, respectively to enable more complex behaviors.In both of these examples, complexity of the internal model increases with information processing needs, but also clearly requires interactions with body size and the environment.Likewise, intelligent behavior in insects is not only determined by changes in the size of the mushroom bodies [7], but also environmental cues and communication with conspecifics.Reliance 1292 (2023) 012019 IOP Publishing doi:10.1088/1757-899X/1292/1/012019 2 on a purely connectionist approach results in an incomplete account of even the simplest forms of cognitive processing.
These examples have obvious lessons for models of intelligent behavior mimicked in computational agents.Yet rather than simply scaling up the size of a nervous system for more complex behaviors, we propose a different principle: computational offloading as a consequence of agent development.Developmental approaches offer a novel approach to agent learning and intelligence more generally.For example, it has been shown that human babies and toddlers extract information from the world much differently than our current methods for training machine learning models [8].Human infants intensively explore their environment, but also intently explore specific objects.This provides a flexible form of experience that can be alternatively wide and deep.Early experience with the environment serves as a scaffold for what will come later on in the individual's lifetime.This has been demonstrated computationally through autotelic agents [9].Autotelic agents are defined by intrinsically-motivated acquisition, and use these internal mechanisms to master skills and tasks in a flexible, open-ended manner.
To better understand how this might lead to general principles, we must first understand the nature of cognitive offloading, particularly with respect to the regulation of agent behaviors.We assume that cognition consists of both an internal model and environment which are regulated together.Building upon this assumption, we then consider the concept of tradeoffs and associated factors in determining the share of internal processing to offloading that characterizes increases in cognitive capacity.A developmental approach with a model of innateness is an excellent way to consider both information processing along with the role of a substrate during acquisition.When viewing cognition from a regulatory perspective, we must consider what drives innate capacity for intelligence, and how this system might balance internal capacity for cognitive processing with offloading and related forms of cognitive externalization.To understand this, we introduce a quantitative theoretical model based on heterochrony.Heterochrony refers to changes in the proportion of agent body size to brain (or internal model) size.This connects the themes of regulatory cognition with embodied agents, and why embodiment plays a critical role in determining offloading capacity in computational agents.

1.1
Defining Cognitive Offloading Risko and Gilbert [10] provide an operational definition of cognitive offloading in humans.Their definition of cognitive offloading is the "use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand".Unpacking this definition provides us with a basis for applying this definition to the world of computational agents.Physical action is represented by "development and externalization".Alter(ation) is represented by "development and externalization".Reduc(tion of)e Cognitive Demand is represented by tradeoffs in response to developmental constraints.
Expanding this definition to non-human animals, Andy Clark [11] provides an example of cognitive offloading using affordances in the environment.Bluefin dolphins utilize vortex rings (hydrodynamic features of the water column) to augment their muscle power during swimming.While bluefin dolphins produce more than ample muscle power to propel themselves through the water column [12], vortex rings are produced by the organisms to provide an external computational mechanism for hunting and refining swimming behavior.Critically, the vortex ring provides a means for environmental feedback, and even allows for complex signals to be stored as affordances.

3
Affordances are environmental features that enable computation or ease of behavior.To fully understand how these affordances are useful for the development of artificially intelligent agents, we must consider the developmental and evolutionary origins of offloading.

1.2
Cognitive Offloading in Development Now we will review two studies that demonstrate the role of development in the emergence of cognitive offloading ability.The first involves a shift towards offloading over the course of development in humans [13].As human babies interact with their environments, they learn how to utilize the affordances available in the world, thus adding cognitive capacity as they mature.A complementary study [14] suggests that reactive behavior coupled with offloading acts to reinforce complex cognition in agents.In both of these cases we can see a clear role for developmental contingency [15], which involves context-specific cycles of interaction dependent on the outcome of past events.Therefore, we can see that change and plasticity are hallmarks of developmental approaches to understanding the emergence of offloading.This implies that a stable biological baseline is also important to include in a computational model of offloading.A computational model of innate processes [16][17][18][19] provides the ability to generate more complex abilities as the developmental process unfolds.

2.1
Offloading as an Intelligent Embodied System A cybernetic-inspired diagram of our embodied intelligent system is shown in Figure 1.Contrary to models of cognition as sequential information processing [20][21], this model highlights the regulatory aspects of naturalistic cognition [22,23].The agent phenotype (an example shown in its embryonic form) is influenced by a model of innateness and direct feedback from a model of the environment.In computational terms, the model of innateness serves as an initial condition for the agent, while the environment is a representation constructed gradually by the agent.Innateness is defined in terms of both body and behavior: the shape, growth condition, and sensor/effector units of the body, along with behavioral primitives such as phototaxis, parameters for stimulus sensitivity, or environmental pre-training.Interactions over developmental time contribute to an internal model of the world that becomes iteratively refined.The model of environment consists of selected features outside of the agent's own development.An agent's operational ensemble of environmental features are in part defined through indirect feedback from affordances that conform to innate capabilities such as grasping, rolling, or sensing light.Affordances also occur in association with their fellow agents and serve to inform meta-representations [24,25], shaping how an agent interacts with its environment.
The sophistication of the agent's internal processing and model of innateness, in addition to the quality of available affordances and environmental model determine how much information is retained in the internal model and how much is distributed to the environment.This will be defined in more concrete biological terms when we discuss developmental tradeoffs in the next section.The sociality of the agent is a secondary factor which we will also revisit in the next section in the form of proximal development, or the advantages of social interactions from the environment.

2.2
Two timescales of naturalistic cognition Development and cognition alike unfold over multiple timescales.There are two timescales of particular concern: morphogenetic and dynamical.The morphogenetic timescale involves physical changes and transformations with respect to the biologically-inspired substrate.Changes in the size, shape, and connectivity of the internal model or agent's body fall into this category.By contrast, dynamical time involves the processing of cognitive information processing, which generally happens in concert with an internal model and environmental structure.Changes in information acquisition and offloading capacity fall into this category.
Temporal changes related to the biologically-inspired components of a developing agent occur in morphogenetic time.Morphogenetic time also includes the emergence of the capacitance properties and architectures involved with information processing via the body's sensors.Morphogenetic time presents a unique challenge to models of cognition.Namely, the rate and extent of growth of the agent's body can determine the capacity of the internal model.With this constraint in hand, intelligent behavior might draw from the richness of the environment.Depending on the presence of patterns in the environment, or the ability to utilize affordances, a particular agent will have an externalization capacity that helps them overcome internal model capacity limitations.

Figure 1. Model of regulation between the agent phenotype and the environment in which the agent is
embedded.Agent phenotype is shown in its developmental state.Affordances are highlighted as an incremental feedback.Arrows refer to the direction of interaction: the model of innateness provides instructions to growth and development of the agent phenotype, while the agent phenotype interacts (bidirectionally) with models of the environment.Our models of the environment also interact with affordances, providing a feedback loop that shapes the model composition.
Dynamical time is the processing of information in the internal model, in concert with interactions between the agent and features of the environment.This includes feedback between the internal model and the external environment.The rate of acquisition is constrained by the unfolding of morphogenetic time: the rate of morphogenetic change determines both the ability of an agent's body to acquire information, as well as the capacity of their internal model.The rate of morphogenetic change is determined by the heterochronic rates of growth and form [26,27] determined as one of an agent's innate factors.

2.3
Developmental tradeoffs For a theoretical understanding of offloading and its relationship to the internal model, we will consider a model of tradeoffs.A developmental tradeoff involves prioritizing one factor at the expense of others given a finite amount of developmental potential.Amongst primate species, metabolic constraints due to food source and nutrition provide a selection mechanism for tradeoffs between body size and number of neurons in the brain [28].In this case, brain growth may be limited.Yet if extending cognition leads to a better diet, it may not negatively impact cognition [29].There also exist functional tradeoffs which prioritize certain cognitive functions over others [30].For example, different cognitive functions such as attention and memory take priority based on the needs of a specific cognitive context [15].Tradeoffs can also include behavioral strategies such as speed and accuracy [31], or higher level phenomena such as robustness, resilience, and performance [32].In the case of speed-accuracy tradeoffs, the balance of contributors involves multiple scales of organization (e.g.physiology, cognitive representation, and social interactions).While these examples may confound us with great complexity in terms of feedback and balances, a typical agent-based implementation might involve only a simple tradeoff that favors either a greater or lesser amount of offloading.Some of these factors include growth and energetics, information, affordances and direct and perception.
Environmental tradeoffs in brain size is critical to our perspective on offloading and development.Evidence from mammals suggests that there is a relationship between brain size, intelligence, and lifespan [33].While increased cognitive demands select for larger brain sizes, factors such as degree of social interactivity can mitigate increases in brain size [34].These findings suggest that while the length of development and enriched environments are important for developing intelligent behavior, the capacity for intelligent behavior can be offloaded in strategically interesting ways.Critically, the process of epigenetics [15] governs the inherent variation of the many factors that can contribute to the tradeoffs proposed here.While we will not consider a formal model of epigenetics here, such a model can help control how this process unfolds in agent development.
One advantage of embodied agents with specified innate components is the ability to control the sensors, affectors, and flow of information.Different configurations of these agent properties provide a means to control and understand information acquisition both dynamically and in terms of contingency.Dynamic information acquisition examines the cumulative effects of exposure to the environment.The accumulation of environmental observations is also based on developmental contingency: repeated observations using the available inputs over time determines how the environment is perceived and (more importantly) utilized.Contingent information processing involves configurations that narrow the possibilities over time.
Taken together, direct perception and affordances can include the body (e.g.peripheral motor system), which can function like a statistical prior and information pre-filter [35].This also helps to integrate the environment with the biologically-inspired components of cognitive regulation, and serves as a conduit of offloading.Stigmergy [36,37] involves cues for collective behaviors, and as such serve as affordances that are particularly relevant to coordinating social and other collective behaviors.In social insects, features in the environment serve as a common landmark or reference point to coordinate collective behavior.Stigmergic mechanisms enable distributed communication across fellow agents, and provide a means for emergent self-organization of more complex behaviors [38,39].Heylighen [37] defines stigmergy as a trace left by an action that stimulates subsequent action, providing an indirect but powerful feedback mechanism for collective behavior.We can also view stigmergy as providing the information needed by individual agents to understand how an object in the environment can be used [40].It is this environmental structure that fosters offloading [41], although this is a necessary but not sufficient condition for promoting offloading behavior.
As they map out their environment, developing agents engage with the social milieu.This results in direct feedback to individual agents from both the environment and other agents.This feedback arises through stigmergic constraints, but stigmergic interactions themselves arise from a shared environment that is easier to structure with meaningful information.More generally, proximal development in the form of emerging social interactions can drive shared cognitive resources, which in turn can increase the potential for further offloading, and even larger internal models (for an example from Primate brains, see [42]).

2.4
Examples of tradeoff based Regulatory Systems The combination of factors that go into a developmental phenotype and their effect on the resulting features of a developmental phenotype is called a tradeoff based regulatory system.An example of such a system for a developmental agentis shown in Figure 2. In the computational agent context, tradeoffs occur at two levels: physiological and cognitive.In our regulatory system example, there is no duality between these two domains.This is supported by the observation of temporal tradeoffs in information processing during psychophysical tasks [43].Yet it is not as simple as an equivalence: one feature of this type of regulatory system is that the developmental mechanisms observed in the organismal physiology can be tied to cognitive development.A similar relationship is relevant to computational agents.To simulate the growth and development of a biological system, a scaling model for growth (energetics) and acquisition (information) will be introduced.The relationship between physiology and information processing is particularly interesting for understanding the potential internal conditions that lead to offloading.An example of how these tradeoffs play out in biological systems can be found in how neural tissue expansion is enabled by increased energetic input, which in turn leads to increased neural processing.This can be demonstrated in the context of Primate fovea [44].The outcomes of different causal factors are partially determined by behavioral strategies such as the explore-exploit tradeoff [45].But it comes down to this: as a developing agent explores their world, neuronal tissue must have the capacity to compute and represent an organism's cognitive needs.If not, the organism must find ways (such as affordances) to encode information in the environment.
Biological systems also enforce constraints through the interactions of these tradeoffs.Earlier, we discussed the role of development constraints as defined by developmental systems theory.But there is another definition of developmental constraints that is more specific to growth in the biological substrate.This model of constraints predicts that brain (neural tissue) size grows proportionally with the body, and this proportionality is expected to be consistent throughout development [46].Depending on environmental demands, neural tissue growth increases via an increased growth rate relative to the dimensions of the body.As the growth rate of neural tissue increases, it becomes limited by energetic demands, which in turn dampens the increasing growth rate.This type of nonlinear positive feedback [47] results from dynamically changing biological/energetic tradeoffs over time.As an aside, this is also consistent with what is known about offloading: using an affordance (such as the internet for reference information) encourages further offloading behavior, and strengthens the reliance on offloading over time via positive feedback [48].

3.
Heterochronic Model of Offloading Now we can map the evidence from neural tissue growth during development to computational agents.We maintain that much of what is said about organisms and neural tissue is also true of computational agents and their internal models (or internal representations).The two main issues involve building a substrate for cognitive capacity and enabling regulation of information acquisition and processing through a regulatory system that includes the environment.Thus, characterizing the potential for substrate growth is important for observing and perhaps even controlling how offloaded information processing proceeds.
Neural architectures are formed and maintained through an interaction between physiological conditions and information processing.The capacity of a neural architecture is related to its size and ability to adapt.Both of these factors are the product of development, and are governed by a scaling relationship known as heterochrony.Put simply, heterochrony involves changes in the timing and duration of growth in development.When growth slows down relative to a baseline, the developmental process yields a less mature and smaller mature phenotype.When growth speeds up, the opposite is true, yielding a more mature and larger phenotype.Since this relationship can be used to determine the potential for internal processing, it also serves as an indicator of whether external processing is needed or will even become predominant.
Our heterochronic model for determining the offloading potential of a given agent is shown in Figure 3. Figure 3A shows that acceleration and deceleration are characterized as changes in the value of parameter k (growth rate).This graph also demonstrates changes in k over developmental time (t).Removing k from its temporal context, Figure 3B shows the relationship between k and offloading potential (o).Generally, this is the rate of internal acquisition, with a negative relationship between k and o (low values of o predict high values of k and vice versa).Finally, Figure 3C shows o as a component of developmental time (t) and plotted against k.Characterized by the function f(t) alongside the relationship shown in Figure 3B, this relationship is predicted to decrease with respect to k until an inflection point is reached for moderate values of o and t.Beyond this point, both o and k have a positive relationship.This captures the notion that offloading can be synergistic with large brains, which becomes particularly pronounced at later stages of development.Why this inflection point occurs in Figure 3C is not clear, but we might expect this scenario to occur due to behavioral plasticity later in development.There is yet another function that allows us to characterize the developmental potential called continuous Gibsonian Information (GI).GI [49] is characterized mathematically by the distribution of sensory structure encountered over time by an agent.The function gi(t) approximates the spatiotemporal structure of the sensory environment in terms of affordances, which can be mapped to parameter o in a positive manner: when the value of gi(t) increases, the value of o also increases, and leads to a greater share of offloading as part of an agent's overall cognitive capacity.

3.1
Heterochrony, Acquisition, and Neural Diversity As information acquisition is the opposite of offloading, the basic version of our model predicts that the acquisition of information and heterochrony have a positive relationship.When the neural architecture gets larger (with a corresponding increase in sensory resolution), the acquisition of information from the environment also increases.Furthermore, the configurations of agent properties can be controlled by developmental growth trends.In agents, the growth of neural processing architectures, the development of sophisticated sensory organs, or size and shape properties of the body can all be determined by the length of time and rate of growth and reconfiguration.
We can also conceptualize development as a discrete set of acquisitions that occur in a specific order.Sequence heterochrony is the relative shifting of these events in time [50].In a biological context, this is called sequence heterochrony [51,52], and characterizes development as a series of interrelated steps.For agent-based development, sequence heterochrony can be thought of as developmental contingency in stages.This modifies our continuous growth model in Figure 3 only a bit, as development out of sequence can be disruptive or even dysfunctional.Indeed, sequence heterochrony can be used to develop a viable model of innateness, and allows us to test the compositional sequence of substrate development.
Given the heterochronic rules of development for our agent, development of a specific agent or population of agents follows a specific developmental trajectory.Developmental trajectories can be characterized by epigenetic landscapes [53,54], where a series of contingencies and the path of least resistance produces the mature phenotype.An epigenetic landscape of computational agents provides a set of configurations that allow us to develop specialized phenotypes specialized for various cognitive tasks and social interactions.

Further Considerations
Rather than asking why internal representations are established, a theory of offloading reliant upon proportional growth and biologically-inspired embodiment recasts the notion of internal models as scalable information processing models that extend into the environment depending upon the constraints of the computational task in question.The scalable component is rooted in agent embodiment and rooted in an model of innateness that provides baseline conditions for the implementation of a wide range of architectures.In the computational context, using a biologically-inspired model allows for offloading capacity to become both predictable and controllable.
Returning to Figure 3, we demonstrate how an innate model of growth and transformation leads to the composition of an agent phenotype that serves to both embody cognition and require a certain degree of cognitive offloading.The potential for offloading, and the continuum mentioned previously, results from the tuning of three parameters that summarize agent-environment interactions.Changes in growth rate (k), offloading potential (o) and its relationship to k over time provide a means to introduce or inhibit offloading under specific conditions represented by locations on the epigenetic landscape.
To consider the distinctions and connections between representationalism and offloading, let us consider interface theory on the one hand and radical embodied cognition on the other.Interface theory suggests that the world is constructed in the mind through natural selection, while the external environment is not at all as we make sense of it [55].This is quite different from radical embodied cognition, which argues that our interactions with the external world (and not mental representation) is the primary driver of our cognitive world [56].We can use these approaches as two extremes in a continuum ranging from zero to absolute reliance on offloading.

Conclusions and Future Directions
We have introduced a model of offloading for computational agents.This is done from a developmental perspective, which allows us to both compose agents in new ways and view the externalization of cognition from an emergent standpoint.In the course of introducing a model of developmental offloading, we introduce a number of concepts.The first is a regulatory model of cognition, and how offloading and affordances play an integral role in this perspective.Second, we stress the importance of developmental tradeoffs in determining how much of cognition gets offloaded during an agent's cognition.Our model stresses the potential for multiple causal factors.Finally, we introduce a heterochronic model for the embodied and cognitive growth and development of cognitive agents.Utilizing the theory of heterochrony in the context of our regulatory approach to cognition allows us to define the nature of our tradeoffs and exactly how our agent is embodied.
Viewing cognition as a regulatory phenomenon allows us to characterize our agent's development as an inherently naturalistic set of interactions, and thus evaluate how offloading in the form of augmenting or compensating cognitive capabilities generated by the internal model (or brain).While developmental resources can be specified in terms of energetic or geometric constraints, they generally act to limit the size and capacity of the agent's internal model.Figure 4 shows the developmental context of these three parameters and two functions: growth rate, or k; time, or t, offloading potential, or o; the codevelopment of internal and external acquisition represented by function f(t); and the distribution of affordances and other sensory features during direct perception represented by the function gi(t).This provides a parameterized counterpart to our boxes and arrows model of a regulatory cognition.In fact, one advantage of our heterochronic perspective is the use of three parameters and two continuous functions that describe the relationship between the development of an agent's body, the proportion of developmental resources, and the development of information acquisition by the agent's internal model.
One aspect of our environmental characterization that deserves further investigation is the distinction between affordances and the concept of stigmergy.An environment contains affordances, which enable a number of opportunities for offloaded cognition, including stigmergy.Since stigmergic behavior represents feedback between offloaded cognitive markers and the internal model, it is interesting from both a regulatory and a direct perception standpoint.Beyond the connections to direct perception, a Markov blanket representation [57] of Figure 1 might be useful in identifying which external features are used in tandem with internal processing.Furthermore, the role of externalized IOP Publishing doi:10.1088/1757-899X/1292/1/01201911 cognition can play a role in encouraging the formation of social groups [34].Energetic constraints on brain size among Hominids have been mitigated by externalized forms of cultural practices such as cooking, which in turn enables greater social interactivity [58].This might be useful to social robotics applications where a heterochronic model might be optimized for agent groups of certain sizes.

Future Directions
To expand the usefulness of this work, we should consider the implementation of our approach using specific computational paradigms.Such paradigms include Partially Observable Markov Decision Processes (POMDPs - [59]) and Reinforcement Learning (RL - [60]).Both POMDPs and RL models are a natural fit to our developmental approach to offloading.Specifically, they are easy means to model our regulatory conception of cognition, which requires both an accounting of partially obscured environmental states (or world states in the parlance of POMDP) as well as feedback between the environment (or policy in the parlance of RL) and an internal model.Genetic Algorithms (GAs - [61]) also provide an acceptable means to implement regulatory cognition, and are actually the preferred means to model the innate features of the agent.Furthermore, POMDP models are compatible with multiagent RL (MARL) models [62], which allows for models representing the development of offloading potential to be extended to populations of interacting agents.In a typical biological setting, constraints restrain more mature phenotypes from exploring the possibilities of novel form [63].This is a guiding principle of our biologically-inspired model of growth and development.Indeed, our model of innateness proposes that the agent's body be constrained in exactly this way.Yet we must return to the difference between this type of constraint and what is posited by developmental systems theory [15,64]; namely, that contingencies are conceptualized as a continual set of interacting cycles between the body, brain, and environment.But this is actually consistent with our regulatory approach to cognition by examining the effect of developmental rules on behavioral plasticity that become manifest later in the development process.Returning to Figure 4, we can see that as the possibility space shrinks in terms of what the body, sensors, and effectors look like (or as parameter k driven by t reaches maturity), the amount of developmental freedom, or number of possible trajectories enabled in part by parameters gi(t) and o, and characterized by function f(t), increases.Stated in terms of our proposed theory: developmental freedom is the increase in the number of information processing alternatives despite the winnowing of physiological configurations [19].Developmental freedom allows for critical periods of neural refinement [65,66], in addition to the incorporation of affordances and other externalized objects into a common cognitive system [67].

Figure 2 .
Figure 2.An example of a tradeoff-based regulatory system.Inputs to the developmental agent (both from feedforward innate and feedback environment) include the innate definition of internal model parameters, rate of brain growth, and the innate definition of the phenotype.Outputs from the agent's development include a mature brain size and internal representation, rate of information acquisition, and stigmergy (cues for collective behaviors).

Figure 3 .
Figure 3.An example of a heterochronic model incorporating growth and acquisition.A: an example of deceleration of internal model:body size growth due to environmental constraints.B: baseline offloading potential in terms of the rate of acquisition given heterochronic trends.C: co-development of internal and external acquisition given the potential for behavioral plasticity.Gray: baseline from Figure 3B.Red: an increase in agent size with the same offloading potential.The icons adjacent to the

Figure 4 .
Figure 4. Developmental time and the increase of trajectories due to developmental freedom.The series of letters at the bottom represent the three parameters and two functions that represent our model of offloading potential and heterochronic growth.Boxes represent periods within the developmental process, denoted underneath by caricatures of developing agent bodies.The right-hand boundary of select boxes features a branching point.Branching structures above the adjacent boxes represent developmental trajectories and their associated contingencies that result.Developmental freedom occurs both later in development and when the number of possible developmental trajectories proliferate.