The Impact of AI on Software Architecture through the Lens of Emergence Theory

Abstract

Emergence theory, rooted in the work of P.W. Anderson (1972) and Ilya Prigogine (1977/1984), explains why AI-driven development will reshape how software is built. This article provides the scientific foundation for that argument, built from the original sources, with no prior knowledge of emergence theory assumed.

1. Introduction: why theoretical grounding is necessary

Architecture patterns come and go. SOA, microservices, serverless, event-driven architecture: each of these paradigms was defended at introduction with pragmatic arguments: better scalability, faster deployments, better separation of responsibilities. Rarely was the question asked: why does this actually work? What underlying principle explains that this is the right direction?

This article posits something fundamental: the evolution of software architecture, from monoliths through SOA and microservices to AI-driven systems, is not an incidental trend but a structural development. To defend that argument against criticism, from architects who see historical cycles, from engineers who consider each transition a pragmatic trade-off rather than a directional force, a solid theoretical foundation is indispensable.

This document provides that foundation. It departs from two of the most cited scientific contributions to the theory of complex systems: P.W. Anderson’s “More Is Different” (Science, 1972) and Ilya Prigogine’s work on dissipative structures and self-organization (Nobel Prize 1977; Order Out of Chaos, 1984). Neither work was written specifically about software, and precisely this makes them more powerful as a foundation: the principles they describe are domain-independent and empirically grounded.

Reading guide

Section 2 introduces emergence as a concept for the uninitiated reader. Section 3 covers Anderson’s core argument and its direct relevance to IT. Section 4 adds Prigogine’s dynamic theory. Section 5 applies both theories to software architecture and AI. Section 6 stress-tests the argument against alternatives.

2. What is emergence? An accessible introduction

Those encountering the word “emergence” for the first time are sometimes deterred by the philosophical connotations. In reality, the concept describes something everyday and recognizable: the phenomenon that systems exhibit properties which, no matter how well you know the individual components, you could not have predicted by studying those components alone.

2.1 An everyday example

Take water. Hydrogen is a flammable gas. Oxygen is a gas that supports combustion. Combine two hydrogen atoms with one oxygen atom, and the result, H₂O, is a liquid that extinguishes fire. No property of hydrogen or oxygen individually predicts liquidity, or the specific heat capacity of water, or the remarkable behavior during freezing where ice is less dense than liquid water. These properties “emerge” at the level of the molecule and are not derivable from the characteristics of the atoms individually.

The same principle applies to ant colonies: no single ant “knows” the colony architecture, yet the colony as a whole builds efficient tunnel systems, manages food distribution, and adapts its behavior at the population level. The collective behavior emerges from simple local interactions.

2.2 Two forms of emergence

In the scientific literature, a distinction is made between two variants:

In this document, “emergence” always refers to the weak variant. This is scientifically uncontested and directly applicable to distributed systems.

2.3 Why emergence is relevant for architecture

Software architecture is about designing systems that exhibit properties we want: resilience, scalability, maintainability, domain coherence. Emergence theory posits that these properties cannot simply be “built into” individual components; they are the result of the structure of interactions between components.

The mechanism has two ingredients: flux and constraints. Flux is directed, sustained flow — energy through a thermodynamic system, information through a network, decisions through an organization. Constraints are the structural limitations on that flow — physical laws, protocols, governance rules, architectural boundaries. Emergence arises when sustained, directed flux operates under constraints, causing the system to restrict its possible behaviors and form new, stable patterns. Without flux: no dynamics, no change. Without constraints: chaos, formlessness. Together: stable structures that organize themselves.

Operational heuristic

Emergence = organized flux under constraint. This is a heuristic, not a law of nature. It works well for a wide range of phenomena, but it has limits — and those limits are addressed honestly in section 6.

This has direct architectural consequences. Anderson provides the theoretical foundation for why new properties arise at each level of scale. Prigogine explains the dynamic mechanism: how systems far from equilibrium generate order, and at what cost. The following sections ground this framework in their original sources.

3. P.W. Anderson: “More Is Different” (1972)

Philip Warren Anderson is one of the most influential physicists of the twentieth century. His article “More Is Different”, published in Science in August 1972, has become one of the most cited philosophy-of-science texts in modern science. Anderson worked at Bell Telephone Laboratories and was later professor at Princeton University; he received the Nobel Prize in Physics in 1977.

3.1 The attack on the constructionist hypothesis

Anderson accepts the reductionist hypothesis: that all phenomena obey the laws of the level below. His central argument is that reductionism does not imply what he calls the “constructionist hypothesis” — the view that knowing the fundamental laws (the laws of elementary particle physics) is sufficient to explain and construct everything. Chemistry is then “merely” applied physics, biology is “merely” applied chemistry, and so on. Anderson rejects this: the ability to reduce everything to fundamental laws does not imply the ability to start from those laws and reconstruct the universe.

Anderson, citing Weisskopf, distinguishes two dimensions of scientific inquiry:

• Intensive research: aimed at discovering fundamental laws (elementary particle physics, fundamental mathematics).
• Extensive research: explaining phenomena from known fundamental laws (solid-state physics, biology, chemistry).

The error, Anderson argues, is to think that extensive research is by definition less fundamental or less creative. This is incorrect. At each level in the scientific hierarchy, properties appear that are fundamentally new and that are not derivable by simple extrapolation from the underlying level.

3.2 The hierarchy of the sciences

Anderson posits that the sciences are hierarchically organized: the entities at each level obey the laws of the level below. But, and this is his central thesis, that does not make the higher level “merely applied” lower science:

Anderson’s thesis: at each transition in this hierarchy, entirely new concepts, laws, and generalizations are necessary, requiring as much creativity and fundamental insight as the level below. Psychology is not “merely” applied biology. Chemistry is not “merely” applied physics.

3.3 Broken symmetry: the mechanism behind emergence

Anderson introduces the concept of “broken symmetry” as the mechanism that explains how new properties arise with increasing scale. A system need not have the symmetry of the laws that govern it; in fact, large systems typically do not.

The most accessible example is the crystal. A crystal is built from atoms that obey the homogeneous symmetry laws of space, but the crystal itself exhibits a particular, specific structure (cubic, hexagonal, and so on) that breaks the underlying symmetry. This structure is not derivable from the laws of the individual atoms; it emerges when the system is large enough and transitions to its lowest-energy state.

Core principle Anderson: broken symmetry

Large systems spontaneously adopt the structure that corresponds to their lowest-energy state, even if that structure has less symmetry than the underlying laws. This structure is an emergent property: it is only visible and definable at the level of the whole, not at the level of the components.

3.4 The N → ∞ argument: scale as qualitative threshold

A central insight of Anderson is that in the transition to large systems (the so-called N → ∞ limit), systems undergo “mathematically sharp, singular phase transitions” to qualitatively new states. This is not a gradual change, but a discontinuous jump in which the system begins to exhibit fundamentally different behavior.

Superconductivity is Anderson’s prime example: all fundamental laws describing superconductivity had been known for thirty years. Yet it took thirty years before the phenomenon could be explained, because the explanation required a new conceptual framework at a higher level. Knowledge of the building blocks was not sufficient. Only analysis of the large system as a whole revealed the new order.

3.5 Analysis works; synthesis does not

One of Anderson’s most practically relevant theses is that the relationship between a system and its parts is intellectually a “one-way street”:

• Analysis (from the whole to the parts) works: by studying a functioning complex system, you can understand it.
• Synthesis (from the parts to the whole) is all but impossible: you cannot reliably build a complex system by assembling all components and expecting the desired system properties to emerge on their own.

This has direct consequences for software design: complex system properties (resilience, scalability, domain coherence) cannot be “built in” by specifying the right components. They require architectural conditions that make the emergence of those properties possible, and they require observation of the system in operation.

4. Ilya Prigogine: dissipative structures and bifurcation

Ilya Prigogine was a Belgian-Russian chemist and physicist who received the 1977 Nobel Prize in Chemistry for his work on dissipative structures: systems far from thermodynamic equilibrium that can spontaneously generate order. His major work “Order Out of Chaos” (1984, with Isabelle Stengers) extends this work to complex systems in biology, economics, and society.

4.1 The problem with classical thermodynamics

Classical thermodynamics describes closed systems that tend toward equilibrium. In a closed system, disorder (entropy) always increases: things fall apart, energy differences equalize. This seems to contradict reality: life, and the technology that humans build, continuously generates new structures and complexity.

Prigogine resolves this paradox by showing that classical thermodynamics applies only to closed systems at or near equilibrium. Systems that are open, that exchange energy and matter with their environment, and that operate far from equilibrium, behave fundamentally differently.

4.2 Dissipative structures: order through energy dissipation

Prigogine introduces the concept of the “dissipative structure”: a system that maintains its internal order by continuously dissipating energy (transferring it to the environment). The order is not static; it is actively maintained by the energy flow through the system.

Examples are plentiful:

• A living cell: maintains internal order by continuously consuming energy (metabolism).
• A weather vortex: a hurricane is a stable structure that extracts energy from the temperature difference between ocean and atmosphere.
• An ecosystem: maintains biodiversity through continuous energy flows via the food chain.

Core principle Prigogine: dissipative structure

A system far from equilibrium can spontaneously adopt an organized, stable structure, provided it continuously exchanges energy with its environment. That order costs energy: stop the energy flow, and the structure collapses. Order is not free.

4.3 Bifurcation points: when systems jump

Prigogine’s bifurcation theory describes what happens when the complexity pressure on a system, the so-called “flux” or energy flow, increases to a critical threshold. At that point, the bifurcation point, the existing structure becomes unstable. The system can no longer remain in its current configuration and faces a choice between two trajectories: chaos, or a new, qualitatively higher order.

This is not a linear transition but a discontinuous jump, a phase transition, in Anderson’s terminology. The new order is qualitatively different from the previous one: it has properties that the previous structure did not and could not have.

Historical examples of bifurcations:

• The transition from unicellular to multicellular organisms: at the moment the environmental complexity and the available biochemical tooling reached a threshold, life bifurcated into a qualitatively new form of organization.
• The invention of the printing press: the complexity pressure on knowledge transfer (growing literacy, reformation, scientific revolution) reached a bifurcation point, and the system reorganized around a new technology.
• The transition from mainframe to client-server architecture in the 1980s and 1990s.

4.4 The energy cost of order, and the threshold value

A critical insight of Prigogine is that order has costs. A dissipative structure is only stable as long as the energy flow is high enough to maintain the order. If the cost of order is higher than the available energy flow, the structure cannot form, regardless of whether it is theoretically optimal.

This explains a recurring pattern in IT history: more granular, more distributed architectures were often theoretically conceivable long before they became practically achievable. SOA was described in concept before the tooling existed to sustain it. Microservices required containerization, CI/CD pipelines, and service mesh infrastructure before teams could bear the operational cost. In each case, the “energy” needed to maintain the architecture — orchestration, observability, deployment automation — exceeded what development teams could sustain until enabling technologies reduced that cost below the critical threshold.

Critical insight

The feasibility of an architecture pattern depends not only on its theoretical merits, but on the ratio between the benefits it provides and the energy costs it demands. Each successive architectural paradigm became viable only when the energy costs dropped below what teams could sustain.

5. Emergence applied to software architecture

5.1 The IT hierarchy

Anderson’s hierarchy is explicitly domain-independent: he places psychology, social sciences, and biology alongside physics in his table. The criteria for his argument are structural: systems composed of components, where at each level new properties appear that are not derivable by extrapolation from the level below.

IT systems are hierarchically layered, and the layers map naturally to Anderson’s framework — but with an important distinction. At each level, there are properties that were designed (the engineering purpose of the abstraction) and properties that genuinely emerge (system behaviors that arise from interactions and that no one planned). Both exist. Conflating them weakens the argument; distinguishing them strengthens it.

The lower levels (Physical, Logical) are largely deterministic: the designed abstractions work as specified, and genuinely surprising emergent behavior is rare. But as the stack rises, a gap opens between what was designed and what the system actually does. At the Application level, no one designs technical debt — it emerges from the accumulation of local decisions under time pressure. At the Service level, no one designs cascading failures — they emerge from interaction patterns between independently deployed components. At the Ecosystem level, no one designs Conway’s Law — it emerges from the structural coupling between organizational and architectural boundaries.

This widening gap is the point. Anderson’s hierarchy is not merely an analogy for the IT stack; it identifies where in the stack emergence becomes architecturally significant. The lower levels are engineering. The upper levels are where emergence theory earns its explanatory weight.

5.2 Broken symmetry and domain coherence

Anderson’s broken symmetry describes how large systems spontaneously adopt the structure that corresponds to their lowest-energy state, even when that structure breaks the symmetry of the underlying laws.

In software architecture, this has a direct analogy in Domain-Driven Design (DDD, Eric Evans 2003): the “right” architectural structure is the structure that best corresponds to the structure of the domain. This is the lowest-friction state: the structure in which the software mirrors the business reality as directly as possible, without unnecessary translation layers.

A monolith breaks the symmetry of the domain: it forces all domain concepts, with their own behavioral rules, data models, and rates of change, into one shared structure. This creates friction: parts that should be independent are coupled through shared state, shared deployment, shared failure modes.

Anderson applied to DDD

Just as a crystal spontaneously adopts its lowest-energy structure, the optimal service structure emerges from thorough domain analysis as the configuration with the lowest architectural friction.

5.3 The bifurcation point: AI as catalyst

AI is itself an emergent property of the hierarchy described in section 5.1. The capabilities of modern AI systems are not derivable from the properties of their individual computational components — exactly as Anderson’s thesis predicts. This makes the dynamic recursive: an emergent phenomenon becomes the catalyst for the next phase transition.

This claim requires scrutiny. In 2022, researchers at Google catalogued more than 137 capabilities in large language models that were fully absent in smaller models but appeared suddenly when models reached a critical scale — without fundamental changes to the architecture. They termed these “emergent abilities.”

The critique followed swiftly. Schaeffer et al. (2023) demonstrated that much of this apparent emergence was a measurement artifact: when evaluated with non-linear metrics — where only an exactly correct answer scores, and anything else scores zero — gradual improvement appears sudden. With continuous metrics, the alleged emergent abilities largely evaporated, showing smooth, predictable improvement instead.

However, Barak (Harvard) argued that for complex multi-step reasoning, emergence may be fundamental: when a task requires ten correct sequential steps, and capability at each step improves incrementally, the probability of completing all steps correctly rises along an exponential curve. There is a real threshold. Krakauer, Krakauer, and Mitchell (Santa Fe Institute, 2025) argued for greater precision: emergence should be reserved for cases where successful task performance is accompanied by evidence of new coarse-grained internal representations — novel abstractions enabling algorithmic compression, not merely performance jumps on benchmarks. Not every unexpected scaling result qualifies. But when a model begins to exhibit qualitatively different behavioral patterns backed by new internal structure, that is a serious candidate for emergence.

Epistemic caution

Whether large language models exhibit genuinely new qualitative capabilities at scale, or whether all apparent emergence is a measurement artifact, remains unresolved. The argument that follows does not depend on settling this debate: it concerns the macro-dynamics of AI-driven software development, not the internal properties of individual models.

Prigogine’s bifurcation theory states that a system bifurcates toward a new order when (a) the complexity pressure is high enough, and (b) the energy cost of the new order falls below a critical threshold. In software, AI is driving both conditions simultaneously: it increases the complexity pressure on existing architectures by enabling faster change cycles, more generated code, and autonomous decision-making at the service level, while simultaneously reducing the energy cost of maintaining more distributed systems through automated orchestration, observability, and deployment.

The framework predicts that the current architectural paradigm will not evolve gradually. It will reach a bifurcation point — a discontinuous jump to a qualitatively different mode of organization. What that new order looks like depends on the constraints in place when the bifurcation occurs: the governance structures, the architectural boundaries, the organizational norms.

6. Limits and alternatives

6.1 Falsifiability

A serious scientific thesis must be capable of failure. The emergence framework applied to software architecture would be falsified if:

• Architecture evolution shows no qualitative transitions — if increased distribution always produces proportionally more of the same properties, never fundamentally new ones.
• All observed “emergent” system properties turn out to be fully derivable from component specifications — if synthesis works as reliably as analysis.
• AI-driven software development is fully explained by deliberate human design, without unplanned self-organizing patterns arising at the system level.

To date, the empirical trends support the framework. But that is not proof. Scientific honesty requires saying so.

6.2 Counter-arguments

Reductionism: “it is just engineering”

The strongest alternative says: there is no emergence, only complexity. Distributed system behavior is the deterministic result of component interactions — fully describable, in principle, by studying those components. What this framework explains well: the mechanism at the micro-level. What it misses: why specific architectural patterns stabilize across independent teams and organizations, and why these patterns are the same regardless of the specific implementation. Reductionism explains components, not structure.

Design and planning: “architects choose these patterns”

Individual systems are designed. Engineers make deliberate architectural decisions. This objection is partly correct: individual architectures are designed. But the macro-dynamics — which patterns become dominant, which standards stabilize, how the ecosystem evolves — are not planned by any single actor. The sum of the designs is itself emergent. A city planner can design individual buildings, but the city as a whole — its culture, its economic patterns, its neighborhood dynamics — emerges from the interactions of all actors. No city planner designed Amsterdam as it is now.

7. Conclusion

The same pattern appears at every level of this analysis: energy flux drives systems far from equilibrium, constraints channel that flux, and the result is new stable structures that no one designed but that appear nonetheless inevitable.

The central thesis of this article — that the evolution of software architecture is most productively understood as an emergent process, driven by the same principles Anderson and Prigogine described in physics and chemistry — survives confrontation with alternatives. Not because it is the only explanation. But because it is the most productive one: it connects mechanism, structure, and dynamics in a single coherent framework.

For software architects, this means something concrete. The question is not which architecture pattern is “correct” in the abstract. The question is: what are the constraints — in the tooling, in the organization, in the governance structures — under which the desired system properties can emerge? And: is the energy cost of maintaining those constraints sustainable?

Emergence does not predict what will happen. It says that new structures will arise — and that the constraints determine which.

8. Literature and sources

Frequently Asked Questions