This work introduces Responsibility Engineering, a practice-grounded method for designing, evaluating, and maintaining behavioral responsibility in AI agents under real-world reuse, pressure, and emotionally variable interaction. The method emerged from sustained pressure-testing of deployed, prompt-driven agents whose behavior appeared stable in development and early testing but degraded over time through repeated interaction. Attempts to correct these failures through prompt optimization alone proved fragile, revealing a structural limitation in how responsibility is specified and held at inference time. Responsibility Engineering formalizes a missing systems layer—the Behavioral Contract Layer—where an agent’s behavioral responsibility is explicitly defined, bounded, and inspected. Within this layer, an agent is defined by its job: a clearly scoped responsibility to hold specific behavior steady when pressure would otherwise cause drift. The function of this layer, termed Judgment Engineering, specifies which aspects of behavior must remain invariant, under which pressures, despite ambiguity. The paper introduces two complementary artifacts produced by this layer: ACT, which defines an agent’s behavioral contract, and BASE, which provides structural mechanisms for holding that contract over time. It further introduces patterns as a means of scoping responsibility, enabling selective stability and sanctioned variation while preventing responsibility collision or unbounded expansion. Using scenario-based, multi-turn evaluation in high-stress interactional contexts, the work demonstrates that ACT and BASE do not eliminate behavioral failure, but render responsibility pressure observable, interpretable, and localizable. Failures surface as component-level tensions rather than opaque context drift, enabling targeted structural refinement instead of ad-hoc prompt rewriting. This contribution is intentionally scoped to inference-time behavioral control and evaluation. It does not propose new model architectures, training methods, or alignment objectives, but provides an applied framework for building AI systems whose behavior remains intelligible, inspectable, and correctable after deployment.
Judy Ossello (Mon,) studied this question.