319 Ensign House 12 St George Wharf London Sw8 2Lu

319 Ensign House 12 St George Wharf London Sw8 2Lu

Lake Arrowhead 2003 Sean Boyle & Julian Pratt

Agent–based working – devices for systemic dialogue


We describe a way of working in the health and social care delivery system that we call agent-based working. It is a response to the challenge in complex adaptive systems – how to engage with the parts and connections while at the same time maintaining an awareness of the whole.

One of the authors has worked with colleagues to explore the application of agent-based computer simulation to health care (Boyle et al., 1998). The other comes from a group that has developed an approach, Whole System Working, that enables local systems to get together to find solutions to their shared concerns (Pratt et al., 1999). Both of these ways of working arise from a mental model of human systems as complex adaptive systems, or more broadly as living systems. Each has limitations. We believe that in some situations their strengths can be complementary when deployed as an integrated approach, agent-based working.

When people seek help from the health and social care system, what happens is influenced by the many choices and decisions made by themselves and by all of the service providers they encounter. The journey of care that each follows is a consequence of the interactions of these choices with the resources available to the system. Each choice is contingent and has knock-on effects elsewhere in the system. And when all of these journeys of care are taken together, it is possible to provide descriptions and measures of the system as a whole. If you look at the aggregate of the whole experience of individuals you see envelopes of journeys of care. If you look at the performance of the service provider organisations you see flows (e.g. referral patterns) and stocks (e.g. numbers of people waiting at a particular point in their journey).

All of us have mental models that we use to make sense of what is going on in this system. These are likely to be partial and incomplete – each of us knows a lot about the bit of the system that we encounter on a daily or weekly basis, but rather little about parts of the system we never interact with.

When the system is not working as one would like, a response may be to try to stand outside and form a picture of what is going on. In doing this, we are attempting to formulate our personal mental model; this can take several forms –images or text, maps or numbers, charts or computer models. Each of these can help to clarify one’s thinking and to share it with others.

Computer modelling in particular is a way of quantifying how systems work. But NHS clinicians and managers have until recently approached computer modelling from quite different perspectives. Managers, encouraged (or driven) by performance targets and indicators such as waiting times, have focused their attention on aggregate measures of the actual system. They have used spreadsheet models or, less often, system dynamics or discrete event simulations to explore the implication of proposed organisational changes. Clinicians, whose focus is on the experience of individual patients, distrust aggregate models because these fail to engage with the highly contingent decisions they make every day. Moreover both clinicians and managers have an inherent distrust of aggregate models, which they recognise to be driven by historical data that they know to be unreliable.

Our experience is that clinicians and managers find common ground over computer modelling when the model engages not just with what is going on at an aggregate level but also with the behaviours that underlie this. This can be achieved through a form of computer modelling called agent-based simulation, which is constructed by specifying within the model all the individual elements in the system and the rules these elements use to make choices and guide decisions.

Agent-based simulations have generally been used in this context of ‘standing outside and forming a picture of what is going on’. A quite different response to the awareness that the system is not working as one would like is to find ways for the people who make up the system to share their own experience and understanding of how things are now, and to explore together ways they might change it. This approach, exemplified by Whole System Working, brings people together for this purpose – often including those generally excluded from strategic thinking. It provides devices for productive conversations.

Constructing an agent-based simulation requires conversations about ‘what is going on around here’ and ‘what are the behaviours that give rise to it’. These conversations have traditionally been carried out in a series of individual interviews carried out by a modelling team acting as external expert. We suggest that constructing such a simulation can provide a device for conversations that include mixed groups of people including managers, front-line providers like clinicians, and people who use the services. When these conversations include ‘how we might change things’, the simulation can support them to explore ‘what would happen if we…’ with far more nuanced understanding about unexpected consequences and knock-on effects than would be possible in unsupported conversations. We call such an approach, in which agent based simulation is used as a device for systemic dialogue, ‘agent-based working’.

We first describe some of the implications of seeing human systems as complex adaptive systems, and one particular approach to acting purposefully in such systems, Whole System Working. We then provide a framework for describing the range of approaches taken to computer modelling of complex systems and give a brief introduction to an agent-based simulation of health care. Finally we describe how we have introduced the process of modelling as a system intervention to several local groups, and their positive response.

2Human systems as complex adaptive systems

Complex systems

The universe is complex. Inevitably, in the real world we can only give attention to parts of it at a time. The value of thinking of things as systems is that we try to give attention simultaneously to the parts, the connections, and the whole. But the necessary simplification involved in limiting what we think about by putting boundaries around any ‘system’ is an arbitrary act which assumes that you can know something about the behaviour of that ‘system’ without knowing about how individual elements outside are connected to those inside.

Another simplification that has been used, often implicitly, is that a system can be divided into parts such that you can build an understanding of the whole by understanding the sets of parts independently of each other. This implies that the parts of the system are connected in such a way that interactions within one part can be understood independently of the rest. Even if the simplification is not warranted it may give useful insights into the way that the system behaves, particularly locally and over short timescales. We describe a system where this assumption holds as complicated and the sets of parts as loosely coupled.

If, on the other hand, we believe interactions within one part of a system cannot be understood independently of the elements in another part, we describe the system as complex. In this situation the system cannot be divided into parts in such a way that it is possible to build an understanding of the whole by understanding the sets of parts independently of each other. A complex system is a whole, woven together. We describe the interactions between the parts in such a system as tightly coupled. Here, ignoring the behaviour of any element will ignore knock-on effects elsewhere in the system.

Thinking of something as a complex system – for example a vortex in a bathtub, a flock of boids (Reynolds, 1987) or an iterated prisoners dilemma computer tournament (Axelrod’s (1984) - we might ask:

  • what are the individual elements or agents that make up the system? This question is not intended to imply that an agent can exist in isolation from other agents, but that it may be useful to give attention to an agent while retaining awareness of its connections, and of the whole system.
  • why do the agents behave as they do? What guiding principles shape their interactions? These guiding principles have been variously described as ‘rules of thumb’ (Pratt et al., 1999), strategies (Axelrod & Cohen, 1999), decision rules (Boyle et al., 1998), heuristics (Gigerenzer et al 1999), or simple rules (Plsek and Wilson, 2001).
  • what is the nature of the communication between the elements?
  • what are the patterns of order (if any) in the behaviour of the system as a whole arising from the interaction of the agents?
Complex adaptive systems

The guiding principles of agents in a complex system may allow them to respond to other agents and the environment. But entirely new possibilities arise when the guiding principles themselves are subject to change.

An agent can be described as adaptive; so can a system. If an agent’s guiding principle changes in a way that is not random but is a response to interaction with other agents (or of other aspects of the environment), then we describe the agent as adaptive. In this situation the system is also adaptive, as its guiding principles have changed.

But the system may also be adaptive although agents are not, ie without existing agents taking on new guiding principles. This can happen either if some existing agents disappear, through for example selection, or if new agents (who may or may not have new rules) appear. Here it is the distribution of guiding principles amongst agents that will have changed.

Adaptation (changes in guiding principles) may or may not lead to changes in behaviour. If there are changes these may be an improvement according to some measure of success, but equally could be a deterioration according to another.

If we think of something as a complex adaptive system – for example an evolutionary computing environment in which new segments of code are enabled to arise by mutation and recombination and are subjected to selection (Axelrod 1997) – we might ask how the agents change their guiding principles. This might include asking:

  • how does novelty arise?
  • what selection mechanisms exist that make some guiding principles more likely to persist than others in that particular situation?

If we think of human systems (those that include people as agents in the system) as complex adaptive systems we will continue to ask these questions and at the same time question their relevance to ‘agents’ that:

  • exist in relationship not in isolation, constantly co-adapting with each other;
  • each play their part in multiple overlapping human systems;
  • are meaning-seeking and continually create and construct their mental models of the world and their guiding principles;
  • exist in language and communicate across time and space;
  • can have some awareness of not just the behaviours of other agents but aspects of the whole including aggregate behaviours and emergent properties

When ‘somebody outside the system’ gives attention to a human system, that person introduces perturbations that reveal that they are inextricably part of the system, ie they are not really outside the system. Even – or in fact particularly – people who see themselves as outside the system and having the role of designer, engineer, boss or dispassionate and objective observer are themselves agents in that system. The same may also apply to people who are not directly concerned with influencing a system but whose behaviour does affect how people in the system behave.

People seek to influence the mental models, guiding principles and behaviours of others, and are themselves changed in the process of co-adaptation. They seek to influence the connections and communications between themselves and others, and between others.

As they do so, they may or may not be aware that in each interaction there is an imbalance of power – one of the agents is more likely to influence the other. The nature of this power is context-specific, multifactorial and constantly changing. Mindell (1995) uses the term ‘rank’ to describe its basis. For example I may feel outranked in a dark alley by somebody who is taller and carrying a stick (physical rank); at work by my boss (rank associated with organisational role ); in a meeting by somebody who is cleverer or a better orator (intellectual rank); among friends by somebody who has learned to be able to cope psychologically or who has access to spiritual strength (psychological or spiritual rank). Mindell describes numerous dimensions of rank, all of which can be thought of as descriptions of the privileges people enjoy in their particular societies, including those of wealth, class, skin colour and gender. A person’s rank on each dimension may be changing or unchanging, but their overall rank in a particular interaction depends on the specific local circumstances, which may change from minute to minute. Rank is co-constructed by the people, as they interact, in each moment.

A health care delivery system includes a range of organisations with formal power structures. People with high ‘role rank’ can influence the behaviours and connections of others in a variety of ways – including hiring and firing, directives, procedures and protocols – that are to a greater or lesser extent coercive. These may be appropriate in a range of situations, particularly those in which it is judged to be necessary to prevent people from doing their worst.

But people deploy many other dimensions of rank and many non-coercive ways of influencing others. These may be appropriate in the variety of situations in which it is judged helpful to enable others to do their best. A few examples are

  • setting an example;
  • giving attention to behaviours they approve of;
  • teaching what they believe to be ‘right’ or ‘useful’; and,
  • promoting dialogue.

When we see clearly discernable patterns of order in human systems (including the multiple overlapping systems that include and are included in the delivery of health and social care) we attribute this to the multiple interactions of hundreds of thousands of agents each day.

When we judge that the health and social care system is performing well, according to some standard of success, we attribute this to its capacity to self-organise. When it is performing poorly, again according to some standard of success, we believe it is still likely to be organising itself, but may be doing so according to a different standard of success – for example if a hospital is performing poorly in terms of providing high-quality and timely care, this may be because people are organising around balancing the budget, or defending professional boundaries, or trying to improve parts of the system.

When we see top-down initiatives having little effect – whether these are the initiatives of a parent, chief executive or prime minister – we are not greatly surprised, given the capacity for self-organisation in human systems. Where we see such initiatives having an effect, desirable or undesirable, we recognise that this is frequently due not to their content but to the symbolic nature of their interventions.

3Whole System Working

Thinking of organisations and networks of organisations as complex adaptive systems provides insights that may be useful to everybody no matter what their role, whether chief executive, manager or front-line worker.

We have, over the last 10 years, chosen to engage with systems from this perspective in a facilitative role. This approach, which is both a way of seeing the world and a wide variety of ways of perturbing it, we have called Whole System Working (Pratt et al 1999). It seeks to influence the connections and communications of people and organisations in the system. We are interested in enabling the system to find its own solutions, not in acting as problem-solvers.

We find that this requires us to give attention to nine aspects, which are summarised in Figure 1. Giving attention simultaneously to all of these enables agents (individuals, teams and organisations) to:

  • explore and reach agreement about the meaning and purpose of this human system;
  • recognise whether they are part of the system or not, and join in if they feel passion to make a difference for this purpose;
  • take responsibility for participating and playing their part;
  • bring many perspectives, each with their own mental models;
  • begin conversations about what is actually going on around here, now. All the work takes place in the here and now too, and the approach involves crafting time and space to allow these conversations to happen;
  • develop a web of connections and communication in what is frequently otherwise an under-connected system;
  • recognise patterns of order arising from these communications;
  • recognise that they constitute a whole as well as parts. A system that knows itself will talk about ‘we’ as well as ‘us’ and ‘them’; and,
  • trust local resourcefulness – trust that, if all the aspects have been given attention, the local system will organise around the shared meaning.

Figure 1: Whole System Working