AI systems increasingly operate as cognitive ecosystems of humans, models, tools, artifacts, verifiers, and institutions, even when they are built, evaluated, and governed primarily through model-centric frames. Drawing on distributed cognition (Hutchins, 1995a; Hollan et al., 2000) and recent work on externalization in LLM agents (Zhou et al., 2026), we argue that cognitive ecosystems provide a valuable unit of analysis for understanding and designing intelligent systems, where the key design target is mediating cognitive structure: the artifacts and arrangements through which an ecosystem carries, transforms, coordinates, preserves, or exposes task-relevant information. A trace, draft, script, research synthesis, protocol log, vision document, or test suite is mediating cognitive structure as soon as it participates in the ecosystem’s reasoning; it need not first become formal or stable. Mediating cognitive structure stabilizes through uptake — through being consumed and relied on by downstream work. It can be strengthened through deliberate acts of validation, specification, or declaration, or through repeated use alone. Embedding into surrounding workflows, toolchains, and routines is how stabilization becomes materially load-bearing. Stabilization is not a terminal or global property of an artifact; it is a scoped, gradient, and reversible status conferred by uptake. Downstream work treats some version, claim, interface, routine, or artifact as a floor for a particular use. A ratchet effect occurs when stabilized structure persists across iterations in ways such that later exploration begins from a higher baseline rather than re-deriving what has already come to rest (Tomasello, 1999; Sterelny, 2012; Henrich, 2015). This unit of analysis surfaces a design opportunity often obscured by model-centric frames: the deliberate engineering of ratchet effects.