A Theory of Agentic Observation introduces a formal framework for understanding observation as a governed computational process rather than a passive act of data acquisition. The paper argues that contemporary agentic systems are architecturally inefficient because tools expose raw substrate state directly to reasoning models, forcing cognition to perform reduction after observation has already occurred. In response, it develops the concept of the Parsimonious Observation Surface (POS): a disciplined observation architecture in which tools emit the minimum semantically sufficient projection required for the next admissible decision while preserving explicit escalation paths into deeper substrate detail. The framework formalizes observation through typed Decision Classes, Well-Formed Sufficiency, Δ-indexed Minimality, Observation Contracts, Projection Calculus, Address Algebra, Escalation Graphs, Projection Normal Forms, Witness-Carrying and Proof-Carrying Projections, policy-governed information flow, and observation-planning mechanisms. Rather than preserving complete semantic state, POS systems preserve the decision-relevant distinctions necessary for coherent action, transforming observation into a first-class architectural layer of agentic computation. By positioning observation as a governable interface between substrate reality and reasoning systems, the paper establishes a foundation for more efficient agent architectures, improved local and edge inference viability, reduced cognitive bandwidth requirements, and a broader theory of observation economics. The work contributes a new perspective at the intersection of agent systems, information theory, epistemic systems architecture, computational governance, and AI infrastructure design.
Adam Ableman Mazurk (Fri,) studied this question.