We propose a multi-agent architecture in which agents operating at fundamentally different abstraction levels — philosophical, mathematical, empirical, and implementational — are connected by a novel intermediary called the Distribution Agent. The framework rests on three related principles. First, token relativity: the epistemic status of any token (abstract or real) is not an intrinsic property of the token itself but a relational one, defined by which agent is observing it. The real token of a pair at one level becomes the abstract token of the pair at the next level down, forming a vertical chain that terminates at purely real tokens accessible to the implementational agent. Second, the separation of horizontal and vertical reasoning: each primary agent operates exclusively within its own abstraction level (horizontal), while the Distribution Agent operates exclusively across levels (vertical), maintaining the vertical chain of token pairs. Third, the 2D reasoning matrix: any chain-of-thought response to a prompt is a trajectory through a two-dimensional matrix whose axes are abstraction level (vertical) and conceptual domain (horizontal). Horizontal movement represents reasoning within a level; vertical movement represents grounding a concept to the next lower level. A complete answer is a trajectory that terminates at the bottom row, where all tokens are purely real. When incoming concepts cannot be matched to existing pairs at any level, the system broadcasts across all levels simultaneously, triggering an insight generation protocol: a new token pair is minted through the collision of reasoning from opposite ends of the abstraction spectrum, extending the matrix with a new column. We argue this mechanism captures a structural analog of human scientific insight and demonstrate its applicability to open problems in fundamental physics.
V. A. Tarasov (Sun,) studied this question.