Mood Modulates, It Doesn't Gate: Orthogonal Task-State and Affective-State Composition in Persistent Agent Systems

Abstract

Many persistent AI agents are organized around a request-response loop in which the language model serves as the primary decision-making runtime. In such systems, affective state—when represented at all—is often coupled directly to behavior: changes in mood can gate actions or alter task selection, producing brittle and caricatured behavior. We present a persistent agent architecture in which affective state and task state are structurally decoupled channels that compose at the behavior level. The architecture separates three concerns: an Execution Engine determines what to do, a Salience Engine determines which behavioral mode applies, and a Modulator Engine determines how the resulting behavior is expressed. The central architectural constraint is that affect has no control path that gates task firing or salience transitions; instead, affect modulates behavioral texture, including generation temperature, verbosity, tone, and prompt framing, without determining whether an action occurs. We describe the implemented architecture, including a VAD/PAD-like affective state with homeostatic decay, a five-state salience machine, and a modulation function mapping affective state to generation parameters. To make the orthogonality claim falsifiable, we define a finite, testable 60-cell reachability matrix spanning the relevant affective and task-state combinations, together with a controlled protocol for measuring modulation effects. The current implementation establishes the architectural orthogonality and reachability properties, and the modulation function is applied end-to-end: the Execution Engine consumes the modulator's temperature, max-token, and prompt-framing outputs in live generation requests, with no control path from affect to task firing.

Keywords