Graph-Driven Development: Reducing Cognitive Load in Large Evolving Systems Through Externalized Understanding

In large, evolving, multi-component systems, a major share of engineering effort is spent on context reconstruction—understanding existing behavior, dependencies, and assumptions—rather than on producing new lines of code. I propose Graph-Driven Development (GDD), a methodology that externalizes developer mental models into a unified graph where code, infrastructure, and runtime artifacts are nodes, and invariants are treated as first-class, versioned, lifecycle-managed entities. The primary goal is cognitive load reduction: enabling both developers and AI agents to operate at the level of system relationships rather than repeatedly reconstructing them through ad-hoc code navigation. This position paper presents the conceptual framework for GDD, positions invariants as a complement to tests for system-level structural properties, and outlines an evaluation methodology. Target domain: distributed systems, legacy codebases, multi-team environments, and AI-assisted development workflows.