Technical Responsibility in Artificial Decision Systems

Agentic AI shifts the governance question from capability to responsibility: systems that plan, use tools, delegate subtasks, and trigger external effects require attribution, auditability, and enforceable authority boundaries. This paper defines technical responsibility as an architectural property, not a mental state. The theory specifies three necessary conditions: total traceability (R1), explicit internal limits (R2), and formal justification capability (R3), formalized through authority, evidence, decision, and disposition. We argue that stochastic or tool-using systems without a deterministic, replayable, non-bypassable governance boundary cannot satisfy these conditions on their own. Evidence-governed architectures can implement them within defined scopes, including accountable non-execution through refusal, blocking, escalation, or pending states. SARA is used as one explicit realization of this framework, while legal and moral responsibility remain with accountable humans and organizations.