Software engineering has never been a neutral technical discipline. While we often discuss it in terms of performance, schemas, and latency, every architectural choice we make is an artifact of a deeper, often unexamined background philosophy. This philosophy is not born in the vacuum of academia but is shaped by the geopolitical dynamics, consumption patterns, and industrial logic of the XX century. For decades, this background has favored a paradigm of Control.
In this paradigm, the machine is a passive tool, and the engineer is the “steersman.” This is what Norbert Wiener (1948) defined as First-Order Cybernetics: the science of control and communication in the animal and the machine. It assumes a clear hierarchy where the human provides the input and the machine provides the deterministic output. It is the philosophy of the factory, the assembly line, and the centralized database. But as we move into the era of Large Language Models and autonomous agents, this “Control” mindset is becoming the primary bottleneck to building reliable systems.
From Steersman to Participant
The shift we are experiencing today is the transition into Second-Order Cybernetics. As Heinz von Foerster (1974) proposed, this is the cybernetics of “observing systems.” In this model, the observer is no longer an external pilot standing at the helm of a static machine. Instead, the engineer and the agent are parts of a recursive loop of mutual influence.
Current developments like Claude Code (2025) represent an early, practical manifestation of this shift. These systems do not merely execute a sequence of hard-coded instructions; they observe their environment (the codebase, the file system, the terminal output) and adjust their internal plans based on the feedback they receive. They exhibit a form of “striving” to maintain a goal, re-aligning their strategy when they encounter a disturbance. This is no longer a relationship of command and obedience, but one of Structural Coupling.
The Logic of Structural Coupling
To understand this new relationship, we have to look toward Niklas Luhmann’s Systems Theory. Luhmann (1984/1995) argued that complex systems, whether social, biological, or digital, are autopoietic. This means thy are self-referential and self-producing. They do not “take in” information from the outside in a literal sense; rather, they are “influenced” by environmental stimuli that trigger changes in their own internal states.
As Dong-hyu Kim (2025) notes in recent research regarding Generative AI-user interactions, the interaction between a human and an LLM is a form of structural coupling. We do not control the internal weights of the model when we prompt it. We provide a stimulus that the model processes according to its own probabilistic logic to maintain its “alignment” with the task.
This realization changes the engineering goal. If we accept that we cannot control an agent at a XX-century level, we stop trying to build rigid guardrails and start building Agentic Alignment Interfaces (AAI). These are homeostatic membranes designed to manage the tension between human intent and machine autonomy.
Homeostatics: The Engineering of Stability
Recognizing the background philosophy allows us to open up possibilities that were previously obscured. In the 20th century, we built tools that functioned only when pushed. In the 21st century, we are building systems with a digital Conatus, concept that comes from Spinoza and means the striving to persevere as viable systems. This is often referred to as “homeostasis” in systems theory, and to me it suggests that this is what the future software will be about: sustaining metastable systems through feedback loops and structural coupling between sub-systems.
While traditional engineering focuses on the “success path,” Homeostatics would focus stabilizing an action-feedback loop, maintaining stability in the face of noise. It is the math behind Karl Friston’s Free Energy Principle, where an agent acts to minimize “surprise” or dissonance between its internal model and the external world.
For a software architect, this means moving away from “vibe coding” and toward the design of systems that can evaluate their own “truth-value” or alignment. Instead of a simple retry loop that fixes a JSON syntax error, a sort of “homeostatic orchestrator” evaluating whether the retrieved information in a RAG pipeline, for instance, is creating “tension” with the user’s ultimate goal. If it is, the system realigns itself autonomously.
The new interface
This is an important shift in engineering. It is not an adversarial move against human agency, but a recognition that we are now designing systems that participate in our social and technical structures as distinct actors. The software we build is no longer just a reflection of our commands; it is a reflection of our ability to align different types of intelligence.
When we ground our engineering in the theory of Luhmann and the cybernetics of Wiener and Ashby, we stop treating AI as a “black box” that we hope will work. We start treating it as a system to be coupled with, managed through alignment, and stabilized through homeostatic design. This is the difference between building a tool and building an ecosystem.
Sources & Further Reading
Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press. Link
Luhmann, N. (1995). Social Systems. Stanford University Press.
Kim, D. (2025). Exploring Generative AI-User Interactions through Self-Programming and Structural Coupling in Luhmann’s Systems Theory. IMR Press. Link
Umpleby, S. A. (2025). Second-Order Cybernetics as a Fundamental Revolution in Science. Link
Ashby, W. R. (1960). Design for a Brain. Chapman & Hall. Archive
Top comments (0)