State Engineering: Transforming Raw Context into Enriched State for Agentic Systems

As large language model based agents evolve from single turn prompt response systems into multi step, tool using workflows, the limitations of prompt centric design have become increasingly apparent. Recent approaches emphasize context engineering, focusing on what information is presented to a model at inference time. While effective for improving isolated reasoning steps, this perspective does not address how information is structured, transformed, and accumulated across an agent’s execution lifecycle. This paper introduces State Engineering, a data science inspired discipline concerned with the deliberate transformation of raw context into enriched, structured, and persistent internal state. We argue that State Engineering plays a role for agents analogous to feature engineering in machine learning, converting unstructured inputs into representations that enable effective reasoning and execution. Unlike context engineering, which is ephemeral and model facing, State Engineering produces multi dimensional state that persists across steps and is consumed by agents, sub agents, and processing workflows. We formalize the concept of agent state, describe its components, and illustrate how state evolution enables more reliable, interpretable, and performant agentic systems.