Skip to main content

Kennedy School Review

Topic / Development and Economic Growth

When Development Isn’t Complicated


“The explanation of the amazingly high standard of rice cultivation in Bali is to be found in Montesquieu’s conclusion that ‘the yield of the soil depends less on its richness than on the degree of freedom enjoyed by those who till it.’[1] – A. Liefrinck, Dutch Colonial Officer, 1887


In the 1970s, Stephen Lansing stumbled upon a labyrinthine network of water temples. Baroque structures strewn across Bali had long captured the imaginations of outsiders, who had typically explained them as centers of worship. Lansing, a young anthropologist, became convinced of a more mystifying purpose: that these ostensibly religious sites were in fact an expansive, decentralized regulatory system governing the island’s irrigation canals.

Each year, Bali faces the prospect of pest outbreaks. Temporarily abstaining from cultivation mitigates that threat by depriving pests of their habitat—and so farmers could, theoretically, all benefit by cooperating to fallow their fields. And yet, if too many fields lie fallow at the same time, the demand for water will later peak simultaneously, overwhelming supply and inducing shortages.

The functioning of the system—to both control pests and ensure enough water for all—then requires discovering some optimal fallowing sequence that minimizes pest damage and maximizes water supply. Remarkably, without any grand blueprint or official directives, Bali’s farmers had done it. The secret to how, Lansing discovered, lay with the temples.

As Lansing and his colleague James Kremer would later explain, the temple networks serve as a distributed cooperating mechanism.[2] Representatives from various subaks (a kind of farmers’ association) congregate annually at temples to coordinate cropping schedules. Coordination is localized; there is no island-wide conference. Two adjacent subaks might stagger their fallowing by a few weeks. Across the system, the subaks’ seasonal decision-making becomes intimately interdependent.

Instead of some deliberate planning by royal engineers, the managerial functions of temple networks are “the product of trial-and-error adjustments by generations of farmers.”[3] Subaks make decentralized decisions, see improvements in yields or not, and use their experience to improve the next time around. What spontaneously emerges is a vast, self-organized, and extraordinarily sophisticated system of water management.

How did an effective regulatory structure emerge absent central planning? Why, as Lansing and Kremer found, is the system so resilient to shocks like droughts and outbreaks? And just how productive is it?

Rather than just a singular story, Bali may offer new views into an old question: how does development happen?


Balinese agriculture typifies a complex adaptive system (CAS), a concept first articulated by physicists and biologists in the mid-20th century. The term evades an exact definition. Researchers studying ant colonies,[4] the immune system,[5] language,[6] the biosphere,[7] cyberspace,[8] and financial markets[9] all describe deep shared structures that suggest these systems operate similarly. John Holland, a leading figure in complexity science, once remarked that the mechanics of economies and embryos are more alike than not.[10]

Generally, CASs are networks of adaptive agents—ones that learn and evolve through feedback—which generate distinct macro-behaviors through their interactions. These system-level behaviors appear unrelated to agent-level behaviors; in a CAS, the ways in which various agents interact produce novel, emergent properties.[11] Neurons are not conscious, but the ways in which they interact produce consciousness.[12] Ants are not particularly smart, but an ant colony behaves intelligently.[13] Autonomously, Balinese farmers are not maximally productive, but cooperation generates a yield-maximizing system.

This distinctive feature—emergence—is the hallmark of any CAS.

Emergence confounds any effort to understand a system by breaking it down into its component parts. In the 19th century, John Stuart Mill observed the inability to locate the properties of water—for example, its wetness—in either hydrogen or oxygen.[14] Wetness emerges from their dynamic interactions.[15]

These emergent properties are the product of self-organization. James Gleick, a science reporter for the New York Times, once remarked how, in the absence of any leader, flocks of birds move “with a seeming intelligence that far transcends the abilities of their members.”[16] Just as a frenetic trading floor self-organizes into a coherent market with global prices, birds self-organize into a coherent flock with global movements. CASs are generative structures free of conscious design.

That a CAS can behave in intelligent ways—entirely unrelated to the intelligence of any individual and without any central commander telling it what to do—in part explains why Lansing was so intrigued. Among Balinese subaks, as with a flock, there were no ingenious rulers and pre-determined strategies. An optimal fallowing pattern across the island emerged from the localized decision-making of interacting farmers.


Key Features of a CAS

CASs are composed of agents, each with a collection of evolving and conditional decision-rules, or strategies, that guide behavior. For example, a simple rule for an animal might be “IF (approaching object in visual field) THEN (flee).”[17] Agents adapt their behavior by adjusting these rules given the outcomes they produce. A more advanced set of rules enables an agent to perform more advanced behaviors within a given environment, like surviving in harsh climates.

Feedback operates on decision-rules, enhancing agent adaptability, and hence performance, over time. This feedback—in the form of better/worse outcomes—allows an agent to assign a rating to any given strategy. The approach of “flee” with “any approaching object” may work to avoid predators, but the animal will soon also go hungry, so its rules are necessarily modified.[18] Adaptation through feedback is a result first of sorting strong and weak strategies and second of trialing new approaches to replace weak ones.[19] As a local environment changes, so will the feedback and so too will the rules: an always-unfolding adaptive process to immediate conditions.

The composition of these agents is hierarchical, whereby agents group together to become new kinds of agents. Chromosomes generate proteins and proteins become cells, which combine to form organs and then organisms, species, and ecosystems.[20] Aggregates behave distinctly, exhibiting new, emergent properties not found within the underlying agents. [21] For example, an organ’s properties cannot be deduced from studying the properties of chromosomes, proteins, or cells.

How a CAS changes, then—how a species evolves, or how an island comes to regulate water use—is ultimately a function of the adaptive abilities of aggregated agents. How well the parts can learn and adapt is the elemental feature of any CAS.[22]


Complex systems are commonly contrasted with complicated ones.

Complicated systems contain many constituent parts, from a typical electrical grid to the flow of patients through a hospital, but behave predictably. The inner workings may feature numerous interactions, often requiring deep expertise to understand and to shape, but the outputs produced are predictably determined by the inputs, whether it’s a flick of a light switch or patient throughput. Improving system performance requires optimizing the performance of constituent parts, typically through centralized control and technical know-how.

This describes some development challenges well.

In April 2016, the world replaced one polio vaccine with another, the largest “switch” of its kind in history.[23] Due to concerns that overlapping vaccines could cause an outbreak, the switch was given only a two-week window. The effort spanned the health systems of 155 countries, requiring extraordinary coordination across health ministries, global health agencies, and NGOs.[24] Representatives of the Pan-American Health Organization reflected that the switch was “without precedents” and realized “astonishing results.”[25] In India alone, synchronization occurred across 27,000 discrete points along the country’s cold chain.[26] The feat showcased the striking ability of experts to manage complicated systems on a global scale with sophisticated planning and technical expertise.

In complicated systems, experts can impose control centrally: they can deploy staff to visit warehouses to ensure compliance with vaccine removal and enforce reporting systems from clinics up to regional agencies.[27] Designing better inputs, like surveillance protocols, yields better outputs, like decreased outbreak risks. The interactions within complicated systems, and therefore the results, can be known, planned, and managed.

The same cannot be said of complex ones.

In complex systems, nonlinear interactions yield effects that aren’t the result of any particular cause but of relationships too complex to isolate; this handicaps an expert’s ability to plan for outcomes. Structure in a CAS is generated by the self-organization of adaptive agents; trying to impose order is typically a recipe for disorder. Understanding and controlling the parts does not imply understanding and controlling the whole.

John Miller, a social scientist at the forefront of complexity research, laments that the social sciences treat most systems as if they were complicated. Reductionism underlies this treatment—as if observing individual tiles gives insight into a mosaic.[28] “The usual proposition,” he writes, “is that by reducing [social] systems to their constituent parts, and fully understanding each part, we will then be able to understand the world. While it sounds obvious, is this really correct?”[29]

As an ambitious social science—one that endeavors not just to understand our social worlds but to manipulate them for the better—development concerns itself with this pivotal distinction. Do most social systems in fact tend toward complexity? Or are most simply complicated? In development policymaking and practice, what are the consequences of confusing the two?


In 1979, the Asian Development Bank (ADB) launched the Bali Irrigation Project to maximize agricultural productivity. Central to the project was a “mandated change to continuous rice cropping for as many subaks as possible.”[30] Each farmer was encouraged to increase individual yields and abandon coordination with neighbors; the Ministry of Agriculture handed out rewards to the highest-producing plots. Studies by foreign consultants predicted that eliminating the rotational cropping schemes would generate tens of thousands more tons of rice per year, which could be sold for export, and which in turn would be used to repay the ADB project loan.[31]

Pest outbreaks and severe water shortages ensued almost immediately. Crop losses reached nearly 100 percent.[32] Consultants resisted calls from farmers to return to the temple-based irrigation system, interpreting the push-back as “religious conservatism and resistance to change.”[33] Project planners “dismissed these occurrences as coincidence.”[34] They encouraged farmers to apply more pesticides and compete harder to maximize productivity. Outbreaks expanded[35] and shortages intensified.[36]

What went wrong?

At the unit-level, farmers could indeed plant more absent cooperative fallowing. Theoretically, productivity increases on each farm would yield an expected proportional increase in island-wide harvests. But this assumed aggregate productivity was a function of individual productivity—an assumption violated by the interdependencies among farmers. A reductionist approach of decomposing the system into its constituent components tells us little about system productivity when the aggregate behaves differently from its parts. Agents may not have been individually yield maximizing, but the system was.

By calculating productivity at the level of individual plots, the consultants missed the whole for the parts.

The political scientist James Scott, reflecting on the cataclysms of Soviet collectivization and Tanzanian villagization, admonished modern agricultural experimentation as too often “incapable of dealing adequately with certain forms of complexity” because it “tends to ignore, or discount, agricultural practices that are not assimilable to its techniques.”[37]

In Bali as elsewhere, a complex reality was made to fit techniques designed for a complicated system.


Complexity suggests a way of understanding how much of the world works—one that sits uneasily with orthodox development theory and practice.

In a complicated world, the whole is the weighted sum of its parts. Total economic output is a function of the output of all firms; literacy rates are a function of how literate each member of a society is. From here, we tend to conclude that to increase economic output, we need only to make individual firms (or farms) more productive.

To echo Miller, while it sounds obvious, is this really correct?

Political economist Owen Barder suggests instead that development phenomena are emergent properties of CAS.[38]

According to this perspective, development describes the capacity of a system to generate desirable emergent properties like productivity, high life expectancy, or low levels of corruption. These properties are a product of self-organizing complexity: for example, the ways in which firms interact with each other and with their social, political, and economic environments generate macroeconomic phenomena, like output as described by GDP. While this “seems obvious,” Barder writes, “it is a surprising departure from the way most economists have normally described development.”[39]

This complexity comes about not by deliberate design, as in a complicated system, but through the adaptive behavior of agents that co-evolve with one another and their environments. For example, a firm makes decisions in response to feedback from other firms and its operating environment, which in turn changes the strategies of other firms and macroeconomic conditions like prices, which in turn shape a firm’s decisions. It’s through this dynamic interaction that agents discover what works—an adaptive process of finding solutions in temporally and spatially specific contexts.

Subaks had no foresight into the design of a yield-maximizing system. But as with any evolutionary process, feedback over time helped to select certain functions that performed well and discard others. Water temples probably won’t work well in California; the mechanism evolved in response to an island’s particular context. Through tiresome trial-and-error, agents searched, selected, and amplified what worked, adjusting as local conditions adjusted. These adaptive strategies gave rise to a complex system with desirable emergent properties.

If emergence at the system level is a function of adaptation at the local level, then complexity ultimately directs the focus of development policymaking and practice toward the adaptive capabilities of local agents. Social safety nets and free mobility, for example, not only support essential capabilities, such as the ability to weather downturns and to trade, but also serve as the foundation for adaptive functions like risk-taking and idea dissemination. Democratic norms are not just normatively desirable but also enable decentralized decision-making and stronger feedback loops as agents trial approaches to local problems.

“At the heart of [CAS] are agents searching for better outcomes,”[40] notes Miller. When development isn’t complicated, solutions rest on them.


After abandoning the ADB’s project in Bali, the Indonesian government searched for solutions to depressed yields. Its answer, the Training & Visit program—modeled after the World Bank’s “technology transfer” programs en vogue during the late 1970s—propagated new agricultural methods that were generated by government research and taught to farmers.[41] Financed by the Bank for over a decade, the program “stressed exclusive dedication to technical information dissemination through a single hierarchical line of command,”[42] from specialists through to field trainers.

Evaluations found that only 25 percent of trainers ever stepped foot into a rice field; training materials could “still be found neatly wrapped in their original plastic containers at provincial training centers” years later; and in some areas, pest outbreaks actually increased.[43]

In 1989, the government pivoted. Rather than forcing “adoption of external information,”[44] an Integrated Pest Management (IPM) program enhanced farmers’ existing capabilities. The government supplied agronomic and ecological concepts in place of directives. It encouraged experimentation with planting times, varieties, and fertilization, and set up select rice fields as “laboratories” for farmers to test concepts. Knowledge spread by way of farmers exchanging experiences with one another.[45] IPM stipulated that the “farmer remains the central manager and independent decision maker.”[46]

Results were profound: IPM farmers experienced higher yields and lower economic variance than their non-IPM counterparts. By the time the program ended in 1999, BPH, a pest infestation that had once ravaged Indonesia, had all but disappeared.[47]

The contrasting programs’ philosophical distinction was not whether the government had a role to play but what role. IPM forwent a technical approach to a complex agricultural system, putting its resources instead behind farmers’ capabilities to search for and trial solutions.

The mindset leap the program illustrated is perhaps even more profound: a kind of trust in the agents themselves. “Let us not fail to note what kind of experimenters these are,” notes Scott, commenting on the strategies of poor farmers. “Their lives and the lives of their families depend directly on the outcomes.”[48]


Grant Tudor is a graduate student at Harvard Business School and the John F. Kennedy School of Government at Harvard University and a recipient of the Zuckerman Fellowship at Harvard’s Center for Public Leadership. Prior, he served as the founder and chief executive of a nonprofit working to increase the impact of products and services designed for the poor.  


Justin Warner is a graduate student at Harvard Business School and the John F. Kennedy School of Government at Harvard University and a recipient of the David M. Rubenstein Fellowship at Harvard’s Center for Public Leadership. He has prior professional experience in management consulting, private equity, and impact investing.


Edited by Anna Mysliwiec

Photo: Rice paddies in Bali // Credit: Skitterphoto from Pexels

[1] F.A. Liefrinck, “Rice Cultivation in Northern Bali,” in Bali: Further Studies in Life, Thought and Ritual, J. van Baal, ed. (The Hague: W. Van Hoeve Publishers, 1969). Excerpted from J. Stephen Lansing and Thérèse A. de Vet, “The Functional Role of Balinese Water Temples: A Response to Critics,” Human Ecology 40, no. 3 (2012).

[2] J. Stephen Lansing and James N. Kremer, “Emergent Properties of Bali Temple Networks: Coadaptation on a Rugged Fitness Landscape,” American Anthropologist 95, no. 1 (1993): 97–114.

[3] Lansing and Kremer, “Emergent Properties of Bali Temple Networks.”

[4] Eric Bonabeau, “Social Insect Colonies as Complex Adaptive Systems,” Ecosystems 1, no. 5 (1998): 437–43.

[5] Max D. Cooper and Matthew N. Alder, “The Evolution of Adaptive Immune Systems,” Cell 124, no. 4 (2006): 815–22.

[6] The “Five Faces” Group et al., “Language Is a Complex Adaptive System: Position Paper,” Language Learning 59, no. s1 (2009): 1–26.

[7] Simon A. Levin, “Ecosystems and the Biosphere as Complex Adaptive Systems,” Ecosystems 1, no. 5 (1998): 431–6.

[8] Paul W. Phister Jr., “Cyberspace: The Ultimate Complex Adaptive System,” The International C2 Journal 4, no. 2 (2010).

[9] C.H. Hommes, “Financial markets as nonlinear adaptive evolutionary systems,” Quantitative Finance 1, no. 1 (2001): 149–67.

[10] John H. Holland, “Complex Adaptive Systems,” Daedalus 121, no. 1 (1992), 17.

[11] John H. Holland, Emergence: From Chaos to Order (Cambridge, MA: Perseus Books, 1998).

[12] L. A. Cacha and R. R. Poznanski, “Genomic instantiation of consciousness in neurons through a biophoton field theory,” Journal of Integrative Neuroscience 13, no. 2 (2014): 253–92.

[13] Kevin Kelly, Out of Control: The Rise of Neo-Biological Civilization (Reading, MA: Addison Wesley, 1994).

[14] John Stuart Mill, A System of Logic Ratiocinative and Inductive (London: Longman, Green and Co, 1843), 243.

[15] John H. Holland, Complexity: A Very Short Introduction (Oxford: Oxford University Press, 2014), 4.

[16] James Gleick, “New Appreciation of the Complexity in a Flock of Birds,” The New York Times, 24 November 1987,

[17] Holland, Complexity, 26.

[18] Holland, Complexity, 26.

[19] John H. Holland, Hidden Order: How Adaptation Builds Complexity (Cambridge, MA: Perseus Books, 1995), 83.

[20] Annalisa Baicchi, Construction Learning as a Complex Adaptive System: Psycholinguistic Evidence from L2 Learners of English (New York: Springer, 2015), 11.

[21] Holland, Emergence.

[22] Holland, “Complex Adaptive Systems,” 17.

[23] Donald G. McNeil Jr., “For Polio Vaccines, a Worldwide Switch to New Version,” The New York Times, 4 April 2010,

[24] “Preparing for the withdrawal of all oral polio vaccines (OPVs): Replacing trivalent OPV (tOPV) with bivalent OPV (bOPV),” World Health Organization, February 2015, accessed 16 February 2018,

[25] Christina Pedreira, Elizabeth Thrush, and Barbara Jauregui, “Systematization of the Introduction of IPV and Switch from tOPV to bOPV in the Americas,” The Journal of Infectious Diseases 2016, suppl. 1 (2017): S76–S85.

[26] P. Haldar and P. Agrawal, “India’s Preparedness for Introduction of IPV and Switch from tOPV to bOPV,” Indian Pediatrics 53, suppl. 1 (2016): S44–S49.

[27] “Toward the end of polio: The vaccine ‘switch’ in the Americas,” Pan-American Health Organization, 29 December 2016, accessed 26 February 2018,

[28] John Miller and Scott E. Page, Complex Adaptive Systems: An Introduction to Computational Models of Social Life (Princeton, NJ: Princeton University Press, 2007), 44.

[29] Miller and Page, Complex Adaptive Systems, 27.

[30] J. Stephen Lansing, Priests and Programmers: Technologies of Power in the Engineered Landscape of Bali, (Princeton, NJ: Princeton University Press, 2007), 124.

[31] Lansing, Priests and Programmers, 124.

[32] John H. Miller, A Crude Look at the Whole: The Science of Complex Systems in Business, Life, and Society (New York: Basic Books, 2015), 173.

[33]  Lansing, Priests and Programmers, 124.

[34] J. Stephen Lansing and Karyn M. Fox, “Niche construction on Bali: the gods of the countryside,” Philosophical Transactions of the Royal Society B: Biological Sciences 366, no. 1566 (2011): 927–34.

[35] Lansing, Priests and Programmers, 124.

[36] Lucas Horst, The Dilemma of Water Division: Considerations and Criteria for Irrigation System Design (Colombo, Sri Lanka: International Irrigation Management Institute, 1998).

[37] James C. Scott, Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed (New Haven: Yale University Press, 1998), 264.

[38] Owen Barder, “The Implications of Complexity for Development,” Center for Global Development, 15 August 2012, accessed 26 February 2018,

[39] Owen Barder, “What Is Development?” Center for Global Development, 16 August 2012, accessed 26 February 2018,

[40] Miller, A Crude Look at the Whole, 19.

[41] Elske van de Fliert, Integrated Pest Management: farmer field schools generate sustainable practices: A case study in Central Java evaluating IPM training (Wageningen: Wageningen Agricultural University, 1998), 16.

[42] Craig Thorburn, “The Rise and Demise of Integrated Pest Management in Rice in Indonesia,” Insects 6, no. 2 (2015): 381–408.

[43] van de Fliert, Integrated Pest Management, 17, 25.

[44] van de Fliert, Integrated Pest Management, 27.

[45] Thorburn, “The Rise and Demise of Integrated Pest Management.”

[46] van de Fliert, Integrated Pest Management, 26.

[47] Thorburn, “The Rise and Demise of Integrated Pest Management.”

[48] Scott, Seeing Like a State, 264.