Context is everything, they say. We say: context is everything which, through its presence in or absence from the system, influences the behavior of an agent.
That definition is so broad it is practically useless. It comes down to: “A thing in the system can be influenced by the rest of the system.” No kidding. For a system observer (or modeler), how much and what ‘rest of the system’ to consider is a very difficult but crucial question though. Not all context is relevant, and defining it too broadly exponentially increases the number of things to keep track of. Include too much, and your system becomes heavy and unwieldy. But include too little, and you will miss out on stuff that matters. Things may not matter at all for a very long time, until suddenly they matter a great deal. Think exposure to Greek banks before and after 2009, or Patient #1 catching a virus in a Wuhan wet market in 2020. Without some carefully considered context built in, a complex system will be little more than an oversimplified, mechanistic machine. The choice of context is always a conundrum, but we can make better informed choices if we understand the different ways context plays a role in system phenomena.
For starters, context determines the degrees of freedom and operational range an agent has within the system. It is the backdrop against which the agents go about their business. Almost all systems need a good bunch of supporting context, without which the agents just don’t function very well. Take cars, for example. Owning a DMC DeLorean in 1985 USA is a very different proposition than taking the same vehicle with you all the way back to 1885. The utility of a car is enormously dependent on the presence of supporting infrastructure. The car is close to useless in a context missing essentials such as a hard surfaced road network, a petroleum industry, service stations, and – very important for a DeLorean – access to spare parts and trained mechanics. Contemporary car use is further facilitated by other, less tangible context such as driver’s license requirements, traffic-related legislation, dedicated law enforcement, and the existence of an insurance industry. The first and most obvious chunk of context provides the rails on which the interesting system bits run. In real life systems, those bits and their rails often develop side by side – a phenomenon known as “co-evolution”. The car-supporting context emerged and developed in parallel with the rise of the car itself. The cars wouldn’t be there without all this context, but the context wouldn’t be there without the cars either.
In the case of natural ecosystems, every carbon-based entity is context for everything else. The set of environmental conditions in a habitat determine the creation and proliferation, selection and adaptation, competition between and extinction of entire agent species – not just the fate of an individual agent. In any ecosystem niche, particularly well suited species thrive and rapidly proliferate. This high pace of reproduction generates variation in the population of individual agents, either through statistical distribution or genetic mutation. The agents best suited to the environment will be most successful at reproducing, leading to generations of agents that are ever more finely adapted to the environment. These forces operate at the level of species just like at the level of individual agents: what works gets gradually reinforced, what doesn’t work gets whittled down. In rich and deep ecosystems, a wide variety of species and many different types of interaction creates stability and resilience in the system. But any change in the context, especially in less diverse systems, can cascade into a new wave of creation, adaptation and destruction causing the system to rebalance in an entirely different state. The arrival of a new predator, overexploitation of a critical resource, extinction of a species due to a parasite, a change in climatic conditions… will reshuffle the contextual conditions, change the rules of ‘success’ and inevitably kick off a new round of adaptation.
This principle of agent-context symbiosis is hard to overlook when the agent visibly relies on or exchanges physical resources with its environment. However when this relationship is intangible rather than physical, a system observer can easily miss it. The problem with analysis is that it’s inherently a reductionist activity. The analyst distills and summarizes, and therefore removes and omits. Analysis seeks to isolate the essential from the unimportant. The analyst is drawn to the image of the Mona Lisa, but may not pay much attention to the landscape behind her. In doing so, it is easy to cut out elements that matter. The business world is rife with examples. Every company in the world is full of hard working, well meaning professionals eager to emulate the “best practices” from industry leaders. And yet, implementing the “Top 3 Amazon team habits” they read in Harvard Business Review doesn’t automatically lead to the same outcome they produce at Amazon. Often that is because the context at Amazon is different from their own organization in a meaningful, but poorly understood way. Some processes, habits or team structures will only produce their desirable effect if they are supported by other processes, tools, computer systems, governance policies or communication habits. Hiring is another example. The hiring company gets a new employee in, but not the context in which he or she produced their track record. Sometimes this is a conscious decision and new employees, especially senior ones, are deliberately brought in to transform the existing company culture. But more generally, overlooking the importance of context producing the track record is the reason the new golden boy or girl doesn’t quite deliver on the initial high hopes. A third example is the use of external consultants. In a successful assignment, consultants don’t make the difference with their understanding of the relevant core best practices for the case. Everybody can do that. What makes the difference is the consultants’ insight in how these principles are at play in the specific context of the client organization. Unlike the physical components that are contextually relevant in a system, which will typically reveal themselves if we stare at the system for a while, it is not so hard to keep missing the intangible, invisible ones no matter how long we look.
For modelers, identifying these relevant intangible components is a major part of the work. In agent-based modeling, half the model is defining and structuring the information exchange between agent and context. (The other half is defining the agent-internal algorithms, taking this data as input and processing it into agent decisions and actions as output.) In machine learning, these modeler-defined algorithms are replaced by a neural network, which uses pattern recognition to predict an output value from the input data. The neural network is a black box version of the algorithms, the modeler gains no insight in any rules or heuristics from this network. Nevertheless the modeler still needs to have a well informed view on which input variables are relevant to produce good output. Agent-based models and machine learning are very successful in producing breathtaking applications these days. That is exciting, but it can lead us to the myopic view that all relevant contextual information is digital or digitizable. That is not the case, some information is relevant but ephemeral. Building a decision engine for the game of go is one thing, building an engine for the political or military decisions of some autocratic head of state quite another. How does one produce reliable signal reflecting ego, or cultural heritage?
So far we’ve only considered context in terms of – physical and informational – conditions surrounding the agent in the here and now. But there is such a thing as temporal context. The individual agents can have memory, and even if they don’t the system as a whole often does. If agents factor in the outcome of earlier decisions in their decision making, we call that learning. The system itself can also accumulate learnings from history and store them in memory. With human beings as the agents, this is essentially how civil society systems emerged. People have – good and bad – experiences interacting with each other, and the accumulated learnings get codified and stored at different hierarchical levels, such as institutions. At a governance level, they are expressed as legislation enhanced with enforcement mechanisms. At an even higher cultural level, they are captured in the norms and values of an ethics. The different historical paths of different agent groups results in different governance and cultural structures, and these differences can be quite dramatic (e.g. religious belief systems) or a little more subtle (e.g. the high sensitivity to monetary inflation in Germany following their experience during the Weimar republic). All in all, this makes the job of our poor system observers and modelers even harder: not only do they need to look widely around in the system, they also need to trace where the system has been throughout its existence…
In summary, perhaps it’s an exaggeration to say that context is everything, but it sure as hell is an awful lot. As always, the system observer or modeler will define or construct the system in function of the research questions they want to analyze. Complex adaptive systems require far more thought going into the context which constitutes the relevant environment of the agents of interest. Agents have a decision engine, applying decision algorithms to information input. Agent decisions can alter their mode of interaction with other agents or with their environment. The decisions and interactions can even lead to the emergence of new agent types and new decision engine algorithms. Unsurprisingly, this means that the selection or design of the right context will be more complex than for regular systems.
 In earlier portraits the landscape’s horizon is almost always aligned with the subject’s neck, but in this case with her eyes – adding to the mystical atmosphere of the painting. Context makes a difference.