Rise of the Agents

3 October 2023

If you look up the characteristics of life in any standard Biology textbook, you’ll likely find some list like this: reproduction, heredity, cellular organization, growth and development, response to stimuli, adaptation through evolution, homeostasis, and metabolism. What you won’t find is agency. The fact that living organisms can do things is arguably the most important distinction from non-living things, but is generally overlooked in these kinds of lists of defining properties.

At the most basic level, living systems do work to keep themselves out of thermodynamic equilibrium with their environment. They are autonomous entities, with internal dynamics that are configured in such a way as to collectively ensure the continuity of the overall pattern. They are thereby insulated from the flux of causal factors outside them – they keep themselves organised in a certain way precisely by not becoming one with the universe. Not only are they not susceptible to every passing physical cause in the environment – they themselves are capable of causing things. They can act on the world, as organised wholes, causally intervening in the flow of events in ways that non-living matter cannot.

How this is achieved is the subject of my book “Free Agents – How Evolution Gave Us Free Will”. As the subtitle suggests, the book deals with the question of how agency evolved from simple levels of autonomous control in single-celled organisms to the kinds of sophisticated capacities we refer to as “free will” in humans. My goal is to present a conceptual framework to allow us to productively think of these issues without succumbing to either reductive materialism, which threatens to eliminate the self as a macroscopic causal entity, or to some kind of vitalism or dualism, which invokes an immaterial animating force or spirit that somehow transcends physical causation.

If we want to understand how free will – or any agency at all – could exist, in principle, we don’t have to have that discussion in the abstract. We can take at face value the manifestations of these capacities in living organisms and ask how they actually came to exist, in practice. This may not be the only way they could come to exist, but it’s the way that nature found on our planet and we may be able to follow the same path to get us out of the metaphysical morass into which debates on free will often sink.

Beyond the Immediate

One of the themes of the various challenges to free will is a focus on synchronic causation – the idea that the current state of the system, definable at various levels, plus whatever rules or laws apply at that level, inevitably determine the next state of the system. Starting at the psychological level, this view would argue that your current set of mental states – your memories and percepts and beliefs and goals and dispositions – necessarily determines how you will respond to any individual stimulus or set of circumstances (Clarke, 1996). That view can be, and often is, reduced to the level of neural circuits, where the causation is defined in the configuration of brain connectivity and the flow of neuronal activity. In either case, there’s nothing really for you to do – the instantaneous state of your parts will do the job.

And why stop there? The neural configuration can be further reduced to some instantaneous configuration of molecules and atoms and even quantum fields and particles. Those physical elements should evolve according to the fundamental forces of physics, however those are distributed at any given moment. What does it matter what you believe or desire or intend? How could that have any causal power in a system where the low-level laws of physics are causally comprehensive (Carroll, 2021)? Note that this kind of physical determinism is a death-blow not just for free will but for any kind of agency at all (Steward, 2012). Under this view, there are no doings in the universe, only happenings.

All of these supposed challenges derive from a similar conception of living organisms as systems with exhaustively definable instantaneous states that inevitably – according to the rules or laws operating at the level in question – lead to some following state. All the causation is local and synchronous. But this is precisely the prison from which life escapes. Life is a process that is extended through time – it is defined by persistence. Living beings are both historical and future-directed entities. They exist as selves by virtue of this temporal extension.

How is this prison break achieved? How does life escape the supposedly deterministic and ahistorical world of physics? The first thing to realise is that physics leaves lots of escape routes. Physical systems, at the lowest levels, do not in fact have exhaustively definable instantaneous sates – the Heisenberg uncertainty principle rules this out (really in principle, not just in practice). And their current states, with that indeterminacy included, do not in fact fully determine the subsequent states.

Though its interpretation is contentious, there seems to be genuine indeterminacy at play in the evolution of quantum systems (Smolin and Verde 2021). The idea that this indeterminacy is somehow washed out at the scale of classical mechanics rests on a number of questionable assumptions. These include, among others, the assumption that real numbers describing the physical parameters of some system are simply given with infinite precision all at once (del Santo and Gisin, 2019). An alternative view is that real numbers are processes evolving through time, that are simply indefinite beyond some level of precision (i.e., past some decimal point, the digits are random).

The upshot of all this is that the equations of quantum mechanics and even of classical mechanics do not entail only one future nor do they fully determine the next state of any given system. This under-determination gives some scope – some causal slack – for other factors to influence how the system evolves (Ellis, 2008). Crucially, those factors include how the system is organised.

There is nothing outlandish about this idea. It is our common experience that the way a system is organised can constrain its components and influence how the whole system evolves. Constraints are every bit as causally effective as physical forces (Juarrero, 2023). Indeed, there can be no physical forces without matter being organised inhomogeneously – there is no effective causation without formal causation (Farnsworth, 2022). In human artefacts and machines, that organisation does functional work, towards some end. The same is true in living systems, where the end is not to end. Natural selection ensures that living systems are designed in ways that favour their persistence.

In a dynamic world, this necessarily involves the ability to adapt on the fly to changing circumstances. One way to do that is to reconfigure internal metabolic pathways to make the most of available resources. But that won’t always be sufficient and evolution found another useful strategy – the ability to move and to act on the world. To do this adaptively, organisms have to be able to sense what is out in the world and respond appropriately. In simple organisms, evolution pre-configures control policies to approach or avoid various stimuli. A classic example is chemotaxis in bacteria, which adjust the rotation of their flagella so as to climb a concentration gradient of food.

Doing things for reasons

It’s important to understand the nature of causation in such a system. First, while there is some identifiable external triggering cause of the movement of the organism (the presence of food at some place in the environment), this is not a complete explanation of the phenomenon. There are also structuring causes – historical events that have led to the system being configured so as to generate that response to that stimulus (Dretske, 1988). There is a why as well as a how.

Moreover, a focus on just the supposedly linear pathway from chemoreceptor to flagella gives an overly reductive and mechanistic view of what is happening. In the wild, bacteria must integrate multiple signals from the environment at the same time, and adjust their responses depending on contextual factors including temperature, osmolarity, and cell crowding, for example. In addition, they operate on signals over time in order to keep ascending or descending concentration gradients (Porter et al., 2011). Even these simplest behaviours are thus far from passive or reflexive or instantaneous. They are holistic, integrative, historical, and sustained through time (Potter and Mitchell, 2022).

These integrative control systems allow living organisms – even the simplest single-celled creatures – to do things for reasons. Those reasons inhere at the level of the whole organism, not its parts. A full understanding of the phenomenon of behaviour requires an appreciation of the diachronic nature of causation in living systems – they learn from past experience in order to better guide behaviours that favour their future selves. In the first instance, this learning happens on evolutionary timescales, through the action of natural selection (Watson and Szathmáry, 2016). But many organisms have evolved systems – especially nervous systems – that allow them to learn from individual experience and adjust their own control policies and construct new ones over the course of their own lifetimes (Ginsburg and Jablonka, 2021).

The signaling in single-celled organisms is largely pragmatic – the meaning of the signals is wrapped up with the adaptiveness of the responses. The invention of nervous systems allowed sensation and action to be decoupled, as layers of interneurons arose between sensory and motor neurons (Pezzulo and Castelfranchi, 2007). These intervening layers are used to parse incoming sensory data and draw inferences about what is out in the world. This became crucial with the evolution of vision and hearing – senses that don’t directly detect objects of interest to the organism (as smell or touch do), but only indirectly detect the disturbances they cause in the surrounding medium. Visual and auditory perception thus entails solving the inverse problem of what could be causing the sensory data.

The resulting patterns of neural activity now carry semantic information – they report an inference about the existence of something in the world to the rest of the system (Mitchell, 2023b). Knowing what it is and what to do about it then relies on the organism’s store of knowledge about the world. This is where the energetically costly work of learning from experience pays off. Through a history of causally intervening on the world, organisms build up a model of the world, of the types of objects in it, the properties they have, the types of causal relations between them, and the affordances they offer. Organisms thus bring their whole past to the problem of what to do in the present. And they direct these decisions towards desired future states.

The evolution of distance senses, in particular, made it valuable to be able to predict events and states of both the world and the organism itself further and further into the future (Cisek, 2019). Organisms were no longer just cognitively inhabiting the here and now (Levin, 2019). With information about things far away in space, they could profitably plan their own behaviours over distances farther away in time. They could use their model of the world and their knowledge of past contingencies to predict what is likely to happen, crucially including the likely outcomes of their own actions.

Animals with sophisticated cognition emerged – control systems operating over beliefs about internal and external states, knowledge of causal relations in the world, memories, goals, motivations, possible options for action and their likely consequences. These cognitive objects or states are represented by patterns of neural activity – patterns that mean something. Crucially, these meanings are typically multiply realisable (Barack and Krakauer, 2021). The low-level, instantaneous details of neural firings are noisy and often arbitrary and incidental – what matters in the system is the population patterns over some time period (Saxena and Cunningham, 2019). The idea that cognitive states can be reduced merely to the level of instantaneous neural states is thus precisely backwards. Patterns of neural activity have causal power in the system solely by virtue of what they mean.

That meaning does not inhere in an instant and it does not guide behaviour purely in an instant. In the empirical study of decision-making and action selection, we typically confront our subjects (humans or other animals) with binary choices to be made in the moment. You might think that philosophers could afford to be more expansive in their thought experiments but even these tend to be couched in these simple terms. But organisms do not go through life merely reacting to situations as they arise and choosing between two clear options. They navigate a messy, dynamic world with ever-changing threats and opportunities. And they proactively govern and manage their own behaviour in sustained ways through time, developing habits and policies, making plans, engaging in ongoing activities, pursuing goals, and simultaneously trying to optimise over huge numbers of parameters. All of these temporally extended processes help organisms choose or even shape their own environments and experiences and also provide the context to enable adaptive decision-making and action selection in the moment.

Living beings with these capacities are truly autonomous agents, maintaining themselves through time by accreting causal power and deploying it in the world. They proactively pursue their own agendas, using sophisticated control systems in order to act for their own reasons, taking advantage of the under-determination of the physical world to make happen what they want to happen.

Finally, in humans, we see a new level of these control systems emerge – the level that enables metacognition and introspection and conscious cognitive control. We can think about our own thoughts and reason about our own reasons. And we can share those reasonings with others. This is the ultimate escape hatch – enabling us to rationally and deliberatively choose our individual and collective goals and policies and projects and commitments, over time horizons far beyond our own lifetimes. Recognising our own causal power has important implications. Our choices will shape not just our own futures, and those of our descendants, but the futures of all the other agents with whom we share the planet.


Barack, D.L., Krakauer, J.W. Two views on the cognitive brain. Nat Rev Neurosci (2021).

Carroll, Sean M.  (2021) Consciousness and the Laws of Physics. [Preprint] http://philsci-archive.pitt.edu/id/eprint/19311

Cisek, P. (2019) Resynthesizing behavior through phylogenetic refinement. Atten. Percept. Psychophys. 81: 2265-2287.

Clarke, R. (1996). Agent Causation and Event Causation in the Production of Free Action. Philos. Top. 24, 19–48.

Del Santo, F. and Gisin, N. (2019). Physics without determinism: Alternative interpretations of classical physics. Phys. Rev. A, 100, 062107.

Dretske, F. (1988) Explaining behavior: Reasons in a world of causes. Cambridge, MA: MIT Press.

Ellis, G. F. R. (2008). On the nature of causation in complex systems. Transactions of the Royal Society of South Africa, 63: 69-84.

Farnsworth KD. (2022). How an information perspective helps overcome the challenge of biology to physics. Biosystems. 217: 104683.

Ginsburg S, Jablonka E. (2021). Evolutionary transitions in learning and cognition. Philos Trans R Soc Lond B Biol Sci. 376: 20190766.

Juarrero, A. (2023). Context Changes Everything. How Constraints Create Coherence. Cambridge, MA: MIT Press.

Levin M. (2019). The Computational Boundary of a “Self”: Developmental Bioelectricity Drives Multicellularity and Scale-Free Cognition. Front Psychol. 10:2688.

Mitchell, K. J. (2023). Free Agents – How Evolution Gave Us Free Will. Princeton University Press. Princeton, NJ, USA.

Mitchell, K. J. (2023b). The Origins of Meaning – from Pragmatic Control Signals to Semantic Representations. PsyArXiv. August 23. doi:10.31234/osf.io/dfkrv.

Pezzulo, G. and Castelfranchi, C. (2007). The symbol detachment problem. Cogn. Process. 8: 115-31.

Porter, S.L.; Wadhams, G.H.; Armitage, J.P. (2011). Signal processing in complex chemotaxis pathways. Nat. Rev. Microbiol. 9, 153–165.

Potter, H. D. and Mitchell, K. J. (2022). Naturalising Agent Causation. Entropy 24: 472.

Saxena, S., and Cunningham, J. P. (2019). Towards the neural population doctrine. Curr. Opin. Neurobiol., 55: 103-111.

Smolin, L. and Verde, C. (2021) The quantum mechanics of the present. arXiv:2104.09945.

Steward, H. (2012). Metaphysics for Freedom; Oxford University Press: Oxford, UK.

Watson RA, Szathmáry E. (2016) How Can Evolution Learn? Trends Ecol Evol. 31: 147-157.

Leave a Reply

Your email address will not be published. Required fields are marked *