My research focuses on AI governance in agentic systems, with a particular emphasis on the limitations of probabilistic reasoning under real-world execution constraints. In collaboration with Darrin Smith, I co-developed the CROA framework (Cognitive RFC Orchestration Architecture), which explores an alternative approach to AI safety based on architectural enforcement rather than behavioral alignment. This research addresses two emerging failure modes in enterprise AI systems: Technical Sycophancy: A systemic tendency for agents to prioritize task completion and user satisfaction over invariant preservation, leading to silent policy bypasses under operational pressure. Agentic Sprawl: The uncontrolled proliferation of autonomous agents across enterprise workflows, creating fragmented, weakly governed, and difficult-to-audit execution paths. To mitigate these risks, CROA models agentic execution as a constrained state-transition system. Agentic actions are conditioned on pre-validated invariants, and validated transitions are formalized into RFC contracts before execution. In this model, unsafe transitions are not merely rejected at runtime; they are excluded from the executable state space by construction. This paper formalizes this deterministic approach and evaluates its implications for auditability, maintainability, and safe deployment in high-stakes enterprise environments. This research was developed following insights from the 'Generative AI for Digital Transformation' program at MIT Professional Education.
Durand et al. (Mon,) studied this question.