Architectural Drift

In the AI Story

The organizational model provided natural resistance to drift through coordination costs. Changing an interface between Team A's component and Team B's component required meetings, negotiations, documentation, testing across both teams. The overhead was high enough that teams only changed interfaces when the benefit clearly exceeded the cost. This friction served an architectural function—it forced teams to think carefully about coupling decisions because the decisions were expensive to reverse. The stability was not optimal (sometimes interfaces should have been changed but weren't, because coordination cost was prohibitive), but it prevented the opposite failure mode: interfaces changing so frequently that the system's structure became unknowable.

AI inverts this dynamic. The designer refactoring her own component interfaces faces no coordination cost—she is both sides of the conversation. Refactoring that would have required two days of meetings and one day of implementation now requires ten minutes of conversation with Claude. The cost is so low that decisions that were previously treated as architectural commitments become tactical experiments. The experimental mindset is valuable for rapid exploration but dangerous for foundational structure. When every coupling decision is revisable at minimal cost, the distinction between foundational and tactical decisions erodes. The system's architectural skeleton becomes fluid, which means unstable, which means its behavior becomes difficult to predict because predictions depend on stable structure.

The Orange Pill's Trivandrum training provides the paradigmatic case. Engineers building with Claude at twenty-fold productivity discovered they could refactor entire subsystems in hours. The freedom was intoxicating—architectural experiments that would have taken weeks could be completed in a morning. But by the end of the week, several engineers reported that they could not reconstruct the reasoning behind certain design decisions because the decisions had been made, revised, revised again, and revised a third time in the flow of rapid iteration. The systems worked, but the knowledge of why they were structured as they were had not been preserved. The coupling structure had become implicit rather than deliberate.

Origin

The concept of drift in engineering has older roots—particularly in Scott Snook's "practical drift" (the gradual divergence between how work is supposed to be done and how it is actually done) and in Diane Vaughan's normalization of deviance (the incremental acceptance of anomalies until they become baseline). Architectural drift applies the same mechanism to system structure: the incremental nature of change makes each step seem reasonable while the cumulative trajectory produces a destination no one would have chosen deliberately. The AI era intensifies drift by compressing the timescale—what would have taken months of organizational drift now occurs in days of individual refactoring.

Key Ideas

Enabled by zero-cost change. When refactoring requires no coordination, the cost of modifying coupling approaches zero. Decisions become experiments, experiments accumulate, structure becomes accidental.

Incremental, not catastrophic. Each coupling modification seems harmless. The cumulative effect is a system whose structure cannot be reconstructed—the designer made decisions faster than she recorded them.

Memory cannot substitute for documentation. The designer believes she will remember the reasoning. Rapid iteration overwhelms memory's capacity. By week's end, coupling decisions exist whose origins are unknown.

Experimental mindset applied inappropriately. Experimentation is valuable for exploration, dangerous for foundational structure. When all coupling becomes experimental, architectural stability erodes—behavior becomes unpredictable.

Requires deliberate recording practice. Simple architectural decision records—listing major coupling choices and reasoning—serve as drift prevention. Externalization transforms implicit decisions into explicit ones, enabling deliberate architecture.