Chapter 1 left autonomy as a driver without a definition. This chapter goes and gets one, and it goes to an unexpected place: living systems are the first systems exhibiting some level of autonomy that we, as humans, have experience and knowledge of. They work as autonomous systems — and, the deck adds as an open question, they evolved to become autonomous.
Two reasons make the detour worth taking. First, the hierarchy of living systems provides examples of many different levels of autonomy, from lower to higher: it is a ready-made catalogue of degrees, which is exactly what a designer needs when the question is “how much autonomy should this component have?”. Second, the evolutionary view potentially sheds light on the role autonomy plays in living systems — on what having more of it buys an organism. The study of living systems may therefore help us understand the many different sorts of autonomy, and their role in artificial systems in general.
Biology is the study of living organisms, or — more generally — of living systems: their structure, function, growth, origin, evolution and distribution. Which immediately raises the question the next section answers: what is a living system?
Read this chapter as terminological groundwork, not as biology for its own sake. Every notion introduced here — boundary, closure, heteronomy, homeostasis, robustness, degrees of freedom — is used again in Chapter 3 when the course finally picks computational autonomy as the defining feature of agents. The biology is what makes that pick defensible rather than arbitrary.
The answer the course adopts comes from Maturana and Varela [Maturana and Varela, 1980; Varela et al., 1974]. Autopoiesis is the ability of a complex system to maintain its own overall coherence, in terms of structure and organisation, through the mutual interactions of its components.
Unpacked, living systems are autopoietic units capable of sustaining themselves, based on an inner network of reactions that generate and regenerate all the system components. All the pertinent processes required have an inner efficient cause — nothing outside the system is doing the producing. Structures, based on a flow of molecules and energy, produce the components that in turn continue to maintain the organised bounded structure that gives rise to those components. Self-reference and self-maintenance are the core notions.
The last clause is the one to memorise, because it is a genuine break with the way engineers usually draw systems: coherent and ordered global system behaviour constrains and governs the behaviour of the individual components, while the component behaviour sustains the global order. This is called circular causality. There is no bottom layer that explains the top, and no top layer that commands the bottom; each is the cause of the other.
Circular causality is what makes autonomy a systemic property rather than a component property. You cannot point at the autonomous part of a cell, because the part that produces the boundary is produced by the boundary. Keep this in mind when Chapter 4 claims that an agent must encapsulate control together with the criterion governing it: that pairing is the computational echo of the same circle.
From this, Maturana and Varela derive their characterisation of autonomous systems: systems that acquire the property of specifying their own rules of behaviour; that do not work as transducers or functions for converting input instructions into output products; and that are the sources of their own activity, which specify their own domains of interaction. In their own words: “the notion of autopoiesis can be described as a characterisation of the mechanisms which endow living systems with the property of being autonomous; autopoiesis is an explication of the autonomy of the living”.
Following Thompson [Thompson, 2010], the boundary is a central element of autonomy: it is a constitutive element of the identity of a system. In a cell, the membrane works as a boundary that both contains processes and components and regulates interaction with the environment. Boundaries are essential for an autonomous system — not incidental packaging, but part of what makes the system a system at all.
Alongside the boundary comes organisational closure [Rosslenbroich, 2014]. Autonomous systems are organisationally closed in the sense that their organisation is characterised by their internal network processes, which recursively depend on each other, and thereby constitute the system as a unit. Closure here is not isolation — matter and energy keep flowing — it is the claim that the organisation, the pattern of mutual dependence, does not reach outside.
Distinguish carefully between boundary (what separates and regulates), closure (the organisation is internally self-referential) and isolation (no exchange — which living systems never have). Confusing closure with isolation is the single most common misreading of autopoiesis, and it leads directly to the mistake corrected in section 6.
The cell is the deck’s worked example of biological autonomy, and it is worth following slowly because every clause does work.
The cell stands out from a molecular soup by actively creating the boundaries that (a) set the cell apart from what is not the cell, and (b) simultaneously regulate its interaction with the environment. The metabolic processes within the cell build and maintain those boundaries — but the metabolic processes themselves are made possible by those boundaries. Overall: a self-sustaining process. The cell therefore emerges as a figure standing out of a chemical background.
And should this process of self-production be interrupted, the cellular components no longer form a unit, gradually diffusing back into a molecular soup — death. Autonomy, in this picture, is not a capability the system has; it is an activity the system continuously performs, and stops being when it stops performing it.
This is the section where biology hands computing a bill. Following Thompson and Rosslenbroich, the deck defines the opposite of an autonomous system:
A heteronymous system is one whose organisation is defined by input–output information flow and external mechanisms of control.
And then names the examples. Traditional computational systems are heteronomous: they have an input layer and an output layer; the inputs are initially assigned from outside the system; output performance is evaluated with respect to an externally-imposed task. Indeed, a Turing Machine typically represents computation by a heteronymous system.
By contrast, an autonomous system is defined by its endogenous, self-organising and self-controlling dynamics. It has input and output — which, alone, do not determine the system. It is the internal self-production process that controls and regulates the system’s interaction with the outside environment.
Organisation defined by input–output information flow and external mechanisms of control. There is an input layer and an output layer; inputs are initially assigned from outside; output performance is evaluated with respect to an externally-imposed task. Traditional computational systems are of this kind, and the Turing Machine is the canonical representation of computation by a heteronomous system.
Defined by its endogenous, self-organising and self-controlling dynamics. It has input and output, but they alone do not determine the system: it is the internal self-production process that controls and regulates the interaction with the outside environment. Its rules of behaviour are specified by itself, and it is the source of its own activity.
Autonomy is not autarchy. Living systems are not independent of their environment; the interchange occurs through the physical boundary. An autonomous system that stopped exchanging would not be more autonomous — it would be dead. The relative, not absolute, character of biological autonomy is stated explicitly in section 10 below.
Be ready to state why the Turing Machine is heteronomous, because it is the cleanest example of the point. Not because it is simple or old, but because its organisation is defined by input–output flow: someone else writes the tape, someone else reads the result, and someone else decides whether the result was the right one. Every classical notion of correctness in computer science presupposes exactly that external judge.
Stated plainly by the deck: living systems are not independent of their environment; the interchange occurs through the physical boundary. Autonomy is a way of managing a relation, not of severing it.
The comparison between plants and animals shows what managing that relation better looks like. Plants exhibit a predominantly open relation to their environment. Animals instead have a more closed form of organisation: the exchange surfaces for metabolism are turned to the inside, special internal organs and internal cavities appear, and exchange surfaces on the outside are reduced.
And here comes the clause that turns a description into a design principle: the loss of a direct environmental relation corresponds to a gain in degrees of freedom. Stimulus–response relationships in animals tend to be less tightly connected; signals can be internally reinforced, compared to other signals, and memorised. What emerges, moving up from plants to animals, is not a rigid but a flexible relation between organism and environment.
Internalisation buys freedom. Every surface you stop exposing to the environment is a coupling you stop being at the mercy of — and the slack you gain is precisely the room in which behaviour can be chosen rather than triggered. This is the biological ancestor of the point Chapter 3 makes computationally: an agent encapsulates control, so nothing outside can invoke it.
Do not read “more closed” as “better” in some absolute sense. The deck presents plants and animals as different arrangements of the system–environment relation, with the animal arrangement trading direct exchange for degrees of freedom. A designer makes the same trade knowingly: a component that internalises everything also loses the cheap, direct, reliable coupling that made it predictable.
Many structures and functions — as well as proteins and genes — show a certain stability in the face of environmental variations and genetic changes: they are resistant, robust, to perturbations, producing relatively invariant outputs [Kitano, 2001].
Robustness is understood as the property that allows a system to maintain its functions against internal and external perturbations and uncertainties. It encompasses a broad range of traits, from macroscopic visible ones down to molecular traits such as the expression level of a gene or the three-dimensional conformation of a protein. It is widely recognised as an inherent property of all biological systems — and clearly connected to their autonomy.
The four principles Kitano lists are, without any translation effort, four software-architecture patterns:
| Strategy | What it does |
|---|---|
| Redundancy | Redundancy of components, to protect against failure of a specific component by providing alternative ways to carry out the function that component performs |
| Feedback circuits | To monitor a system function so as to regulate it |
| Modularity | Encapsulation of functions, for robustness and evolvability |
| Layering | In hierarchical systems, to enhance control and robustness |
Homeostasis is the ability of a system to regulate its internal conditions to keep some or several functions stable. Properties such as temperature or blood composition in animals are homeostatic variables — also called homeostatically regulated variables. The mechanism separates internal and external environments, keeping the internal environment relatively stable with respect to external perturbations.
The deck marks the consequence with an exclamation: homeostatic variables are an obvious manifestation of autonomous behaviour. They are also, it adds, relevant to the biological notion of intelligence — which is exactly where Chapter 5 will pick the thread up, with homeostasis as the very first form of intelligent behaviour on Earth.
Homeostasis is the single notion that links the two drivers. It is presented here as a manifestation of autonomy, and in Chapter 5 as the substrate of the earliest intelligence — a collection of how-to rules relentlessly executed with no explicit directions. Being able to say that in one sentence answers the “what is the conceptual bridge between intelligence and autonomy” question that Chapter 1 left open.
A short section carrying a large idea. Living entities establish their own cycles in time: metabolism, rest–activity cycle, development, reproduction. This involves all biochemical, cellular and organic processes. Reaction rates and frequencies are endogenous, leading to autonomous cycles — which only later synchronise with external cycles.
The ordering matters. The organism does not receive its rhythm from the day and the season and then internalise it; it generates its own rhythm and afterwards brings it into register with the world. An autonomous system, in other words, owns its clock.
Chapter 1 observed that distribution destroys the notion of system time. Time autonomy says something complementary and stronger: an autonomous component does not merely lack a global clock, it has its own and synchronises on its own terms. The two observations together are why the agent abstraction of Chapter 3 has to encapsulate the thread of control rather than a schedule imposed from outside.
Evolutionary biology studies how evolutionary processes produced the diversity of life on Earth — that is, how biological systems evolved over the ages [Gould, 2002]. Its view therefore includes not just how organisms are made and how they work, but mainly how they evolved towards their current form. So: what drives the evolution of living systems?
Rosslenbroich’s answer, adopted by the deck, is increasing autonomy [Rosslenbroich, 2014]. It is defined as an evolutionary shift in the system–environment relationship, with two components that should be kept apart:
| Component | Definition |
|---|---|
| Interactive autonomy | The direct influences of the environment on the respective individual systems are gradually reduced |
| Constitutive autonomy | Stability and flexibility of self-referential, intrinsic functions within the systems are generated and enhanced |
With respect to the environment, the autonomy of living systems is relative: organisms retain numerous interconnections with, and dependencies on, the external environment. What they undergo is a relative emancipation from environmental fluctuations, gaining self-determination and flexibility of behaviour.
A set of resources can be involved to change autonomous capacities. These five are the deck’s answer to “what would you actually change, if you wanted a system to be more autonomous?”
Three claims close the argument. Autonomy is an essential trait of living systems. Biological systems evolve towards increasing degrees of autonomy. Autonomy is therefore an evolutionary pattern — many additional functions resulting from evolutionary processes tend to improve autonomy. The deck itself keeps the last step as a question rather than a doctrine: how can autonomy help in understanding evolution, and is it really a pattern of evolution?
Everything above converges on one definition, which is the single most quotable item in the deck:
General autonomy [Rosslenbroich, 2014]. Living systems are autonomous in the sense that they maintain themselves in form and function within time, and achieve a self-determined flexibility.
Spelled out in six points, they:
Two distinctions from this chapter are worth memorising verbatim, because they are short and load-bearing: heteronomous vs autonomous (organisation defined by input–output flow and external control vs endogenous self-organising and self-controlling dynamics), and interactive vs constitutive autonomy (environmental influence reduced vs internal self-referential functions strengthened). The six-point definition is best remembered through its structure — inner network, boundary, own rules, part–whole interdependence, own time, robustness — rather than word by word.
The deck ends not with conclusions but with two conceptual exercises addressed to the reader. They are the reason the biology was worth doing, and they are open questions rather than doctrine.
Conceptual Exercise 1. Try matching the properties of biological systems with those of complex computational systems, complex artificial systems, complex socio-technical systems and complex intelligent systems.
Conceptual Exercise 2. Try comparing the evolutionary path of biological systems with the potential evolution of the same four families of complex systems.
The widget above is one attempt at the first exercise, and it should be read as such. What the course itself does with these lessons is narrower and more decisive: of the six points, it will keep the third one — specifying their own rules of behaviour — and give it a precise computational content in the next chapter. An agent encapsulates control together with the criterion for governing it. Everything else in this chapter is the argument that such a choice captures something real rather than a metaphor.
The second exercise is the more provocative one. If autonomy is an evolutionary pattern for living systems — if evolution proceeds through systems of increasing autonomy — then asking whether artificial systems are on the same trajectory is not science fiction but a question about the last seventy years of programming languages. Chapter 3 answers exactly that, with Odell’s ladder from monolithic programs to agents.
The ability of a complex system to maintain its own overall coherence, in terms of structure and organisation, through the mutual interactions of its components. Living systems are autopoietic units capable of sustaining themselves, based on an inner network of reactions that generate and regenerate all the system components, where all pertinent processes have an inner efficient cause.
Coherent and ordered global system behaviour constrains and governs the behaviour of the individual components, while the component behaviour sustains the global order. Neither level is prior. It is central because it makes autonomy a systemic property: in a cell, metabolic processes build the boundary and the boundary makes those metabolic processes possible, so no part can be pointed at as “the autonomous one”.
Because a heteronomous system is one whose organisation is defined by input–output information flow and external mechanisms of control, and that is exactly the Turing Machine setting: there is an input layer and an output layer, the inputs are initially assigned from outside the system, and output performance is evaluated with respect to an externally-imposed task. Traditional computational systems in general are heteronomous in this sense.
It is a constitutive element of the identity of the system, not mere packaging. In a cell the membrane both contains processes and components and regulates interaction with the environment. Boundaries are essential for an autonomous system; the interchange with the environment occurs precisely through the boundary.
It means the system’s organisation is characterised by internal network processes which recursively depend on each other, thus constituting the system as a unit. It does not mean isolation: autonomy is not autarchy, living systems are not independent of their environment, and the interchange occurs through the physical boundary. Closure is about the organisation, not about the flows.
Plants have a predominantly open relation to the environment; animals a more closed organisation, with metabolic exchange surfaces turned inside, internal organs and cavities appearing, and external exchange surfaces reduced. The key clause: the loss of a direct environmental relation corresponds to a gain in degrees of freedom — stimulus–response links become less tightly connected, signals can be internally reinforced, compared and memorised, and a flexible rather than rigid organism–environment relation emerges.
Redundancy of components (alternative ways to carry out a function); feedback circuits (monitor a function so as to regulate it); modularity (encapsulation of functions, for robustness and evolvability); layering in hierarchical systems (to enhance control and robustness). All four are recognisable software-architecture patterns, which is part of the point.
The ability of a system to regulate its internal conditions to keep some or several functions stable — temperature or blood composition being homeostatic variables — separating internal and external environments and keeping the internal one relatively stable against external perturbations. It is flagged as an obvious manifestation of autonomous behaviour and as relevant to the biological notion of intelligence, which Chapter 5 develops: it is the bridge between the two drivers.
Living entities establish their own cycles in time — metabolism, rest–activity cycle, development, reproduction — involving all biochemical, cellular and organic processes. Reaction rates and frequencies are endogenous, leading to autonomous cycles which only later synchronise with external cycles. The system owns its clock, and aligns it afterwards.
Both are components of Rosslenbroich’s increasing autonomy, defined as an evolutionary shift in the system–environment relationship. Interactive autonomy: the direct influences of the environment on the individual system are gradually reduced. Constitutive autonomy: the stability and flexibility of self-referential, intrinsic functions within the system are generated and enhanced. Reducing what the world does to you, versus strengthening what you do for yourself.
(1) Changes in spatial separation from the environment (membrane, epithelia, epidermis, cuticula, shells); (2) changes in homeostatic capacities and robustness (conformer/regulator); (3) internalisation of structures or functions; (4) increase in body size (reduction of the surface/volume ratio); (5) changes in the flexibility within the environment, including physiological flexibility, movement capacity and behavioural flexibility.
Living systems are autonomous in that they maintain themselves in form and function within time and achieve a self-determined flexibility, because they: (1) generate, maintain and regulate an inner network of interdependent, energy-consuming processes which in turn generate and maintain the system; (2) establish a boundary and actively regulate exchange with the environment; (3) specify their own rules of behaviour and react to external stimuli in a self-determined way; (4) establish an interdependence between system and parts, including differentiation in subsystems; (5) establish a time autonomy; (6) maintain phenotypic stability, that is robustness, against environmental, internal and genetic perturbations.