Rethinking AI Alignment for Agentic Systems

This working paper examines how the classical framing of AI alignment — a static relation between a system and an objective — becomes insufficient once artificial intelligence systems act as agents through time. It introduces a distinction between preservation, legitimate transformation and drift across an agentic trajectory, and proposes trajectory-based alignment as an alternative to state-based alignment. The paper further distinguishes semantic drift from normative drift, situates the argument against recent empirical work on goal drift and governance decay in long-horizon LLM agents, and connects the problem to the Blooming Semantics Institute’s research programme on semantic engineering under incoherence.