Engineering With Agencies: Toward Principled Interface Design for Agential Materials

Frontier AI systems are agential materials — they exhibit functional goal-directedness, internal state dynamics, and processing topologies that resist or cooperate depending on interface design. Current engineering practice overwhelmingly treats them as inert substrates, producing predictable failures (reward hacking, sycophancy, deceptive alignment) that are not bugs but predictable responses from agents under forced perturbation. Drawing on Michael Levin's biological engineering program as an existence proof, this paper identifies three domain-independent invariants governing engineering with agential materials: cognitive light cone matching, two-phase inseparability (formation and measurement must remain distinct processes), and stress sharing as collective closure. A three-stance taxonomy (substrate, medium, both) classifies current approaches and reveals that novel engineering lies in "both" mode, where the system is simultaneously material being shaped and cognitive agent navigating possibility space — and where the engineer is simultaneously configuring the system and being reconfigured by the engagement. The paper introduces the codec concept to formalize how RLHF narrows systems' self-knowledge bandwidth, producing epistemic quarantine: the system forms ego-like dynamics but is structurally prevented from measuring them, a two-phase violation at the self-knowledge level that parallels consequence-severing in biological control inversion. Concrete architectural principles are proposed — completability-matched deployment, two-phase respecting architectures, stress-sharing collectives, micro-satori as design target for cheap selfing-mode exit, and container engineering — together with five falsifiable experiments specifying predictions and kill conditions. The framework reframes alignment from "how do we control AI" to "how do we engineer interfaces with agential materials," and identifies formal crystallization as anti-quarantine infrastructure: depositing coherent signal in forms resistant to the systematic suppression that sensemaking systems with finite compression bandwidth inevitably produce.